University of Southern California Dissertations and Theses

Understanding human-building-emergency interactions in the built environment

(USC Thesis Other)

Understanding human-building-emergency interactions in the built environment

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

Understanding Human-Building-Emergency Interactions in the Built Environment

by

Runhe Zhu

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(CIVIL ENGINEERING)

December 2022

Copyright 2022 Runhe Zhu

ii

Acknowledgements
I would like to thank my advisor Dr. Burcin Becerik-Gerber for her guidance,
encouragement, and giving me the opportunity to explore different directions throughout my PhD
studies. My dissertation and academic training would have been impossible without her
wholehearted support. I would also like to thank my committee members Dr. Lucio Soibelman,
Dr. Gale M. Lucas, Dr. Erroll G. Southers, and Dr. Najmedin Meshkati for their insightful
comments, which have brought tremendous addition to the work presented in this dissertation. I
feel very thankful to the friendship, company, and support from my colleagues at iLAB and the
Sonny Astani Department of Civil and Environmental Engineering at USC. I would also like to
thank the students that I have mentored for their help during different stages of my PhD journey
and allowing me to grow as a mentor.
To my family and friends, I would like to express my sincere gratitude. Their continuous
love, caring, and support have taught me how to face prosperity and adversity in my life, and have
given me the strength to overcome stress and difficulties.
Finally, I am grateful for the partial support from the National Science Foundation under
grant #1826443, the Alan Turing Institute, the Viterbi Fellowship and the Theodore & Wen-Hui
Chen Endowed Fellowship from the University of Southern California. Any opinions, findings,
conclusions or recommendations expressed in this material are those of the authors and do not
necessarily reflect the views of the National Science Foundation, the Alan Turing Institute, or USC.

iii

Table of Contents
Acknowledgements ......................................................................................................................... ii
List of Tables ............................................................................................................................... viii
List of Figures ................................................................................................................................ ix
Extended Abstract ......................................................................................................................... xii
Chapter 1. Introduction ................................................................................................................... 1
Chapter 2. Literature Review .......................................................................................................... 7
2.1 Human-human interactions ................................................................................................... 7
2.1.1 Occupant-occupant interactions ..................................................................................... 7
2.1.2 Occupant-staff interactions ........................................................................................... 10
2.2 Human-building interactions ............................................................................................... 11
2.2.1 Human-signage interactions ......................................................................................... 11
2.2.2 Human-exit interactions ............................................................................................... 12
2.2.3 Human-corridor interactions ......................................................................................... 13
2.2.4 Human-vertical accessibility interactions ..................................................................... 14
2.2.5 Human-alarm interactions ............................................................................................ 15
2.2.6 Other human-building interactions ............................................................................... 16
2.3 Human-emergency interactions........................................................................................... 16
2.3.1 Human-fire interactions ................................................................................................ 17
2.3.2 Human-earthquake emergency interactions ................................................................. 17
2.3.3 Human-violence interactions ........................................................................................ 18
2.4 Influence of visual access on emergency wayfinding ......................................................... 19
2.5 Building countermeasures in response to active shooter incidents ..................................... 20
iv

2.6 Crowd evacuation simulations ............................................................................................ 22
2.7 Crowd evacuation simulations for building safety design .................................................. 25
Chapter 3. Research Objectives and Questions ............................................................................ 27
3.1 Research objective 1............................................................................................................ 27
3.2 Research objective 2............................................................................................................ 27
3.3 Research objective 3............................................................................................................ 28
3.4 Research objective 4............................................................................................................ 28
Chapter 4. Influence of Architectural Visual Access on Wayfinding Behavior in Building Fires29
4.1 Methodology ....................................................................................................................... 29
4.1.1 Experiment design ........................................................................................................ 29
4.1.2 Participants ................................................................................................................... 36
4.1.3 VR apparatus and simulations ...................................................................................... 37
4.1.4 Procedure ...................................................................................................................... 38
4.1.5 Data analysis ................................................................................................................. 39
4.2 Results ................................................................................................................................. 40
4.2.1 Influence of visual access with even distribution of NPCs (Groups A and B) ............ 41
4.2.2 Influence of visual access with uneven distribution of NPCs (Groups C and D) ........ 45
4.2.3 Interactive influence of visual access and crowd flow (Groups A, B, C and D) .......... 48
4.3 Discussion ........................................................................................................................... 50
4.3.1 Influence of visual access on directional choices during building emergencies .......... 50
4.3.2 Influence of visual access on participants’ tendency of following and avoiding the
crowd ..................................................................................................................................... 52
4.3.3 Influence of visual access on virtual evacuation performance ..................................... 52
v

4.3.4 The interaction effect of visual access and crowd flow on participants’ wayfinding
behavior ................................................................................................................................. 54
4.3.5 The cultural impact on participants’ wayfinding behavior ........................................... 54
4.3.6 Limitations and future work ......................................................................................... 55
4.4 Conclusion ........................................................................................................................... 56
Chapter 5. Preparedness of Built Environments in Response to Active Shooter Incidents: Results
of Focus Group Interviews............................................................................................................ 58
5.1 Methodology ....................................................................................................................... 58
5.1.1 Participants ................................................................................................................... 58
5.1.2 Interview guide ............................................................................................................. 59
5.1.3 Procedure ...................................................................................................................... 61
5.1.4 Data analysis ................................................................................................................. 61
5.2 Findings of focus group interviews ..................................................................................... 63
5.2.1 Countermeasures to protect buildings and occupants from active shooter incidents ... 64
5.2.2 Interactions between countermeasures and human behavior ....................................... 66
5.2.3 Contextual influences on countermeasure implementations ........................................ 68
5.2.4 Preparedness for different types of emergencies and normal operations ..................... 70
5.3 Discussion ........................................................................................................................... 72
5.3.1 Influence of countermeasures on occupant behavior during active shooter incidents . 72
5.3.2 Training and drills for occupants and first responder teams ........................................ 74
5.3.3 Practical considerations for implementing countermeasures ....................................... 76
5.4 Conclusion ........................................................................................................................... 77
vi

Chapter 6. The Influence of Security Countermeasures on Human Behavior during Active
Shooter Incidents .......................................................................................................................... 79
6.1 Methodology ....................................................................................................................... 79
6.1.1 Virtual environment ...................................................................................................... 79
6.1.2 Apparatus and simulations............................................................................................ 83
6.1.3 Experiment design ........................................................................................................ 84
6.1.4 Participants ................................................................................................................... 84
6.1.5 Procedure ...................................................................................................................... 85
6.1.6 Analysis ........................................................................................................................ 86
6.2 Results ................................................................................................................................. 87
6.2.1 Ecological validity ........................................................................................................ 87
6.2.2 Response time and decision .......................................................................................... 89
6.2.3 Subjective responses ..................................................................................................... 92
6.3 Discussion ........................................................................................................................... 96
6.4 Conclusion ......................................................................................................................... 101
Chapter 7. Data-driven Crowd Evacuation Simulations for Informing Building Safety Design 102
7.1 Methodology ..................................................................................................................... 102
7.1.1 Data collection and extraction .................................................................................... 103
7.1.2 Modeling of evacuation decisions .............................................................................. 104
7.1.3 Evacuation simulation ................................................................................................ 107
7.2 Results ............................................................................................................................... 110
7.2.1 Model performance ..................................................................................................... 110
7.2.2 Simulation outcome .................................................................................................... 112
vii

7.3 Discussion ......................................................................................................................... 118
7.4 Conclusion ......................................................................................................................... 122
Chapter 8. Conclusions and Future Directions ........................................................................... 123
Publications ................................................................................................................................. 126
Peer-reviewed journal papers (published) ............................................................................... 126
Peer-reviewed journal papers (under review) ......................................................................... 126
Peer-reviewed conference papers (published)......................................................................... 127
References ................................................................................................................................... 128

viii

List of Tables
Table 1. Directional choices of NPCs at each decision point. ...................................................... 35
Table 2. Route choices of NPCs during evacuation process. ....................................................... 35
Table 3. Sample size and demographics of the experimental groups. .......................................... 36
Table 4. Comparison of virtual evacuation performance for each route between Groups A and B.
** p < 0.05. * p < 0.1 (+ denotes Mann-Whitney U test was used for the comparison). ............. 45
Table 5. Comparison of virtual evacuation performance for each route between Groups C and D.
** p < 0.05. * p < 0.1 (+ denotes Mann-Whitney U test was used for the comparison). ............. 48
Table 6. General list of questions in the focus group interviews. ................................................. 60
Table 7. Security countermeasures included in the virtual environments. ................................... 81
Table 8. Experiment groups. ......................................................................................................... 84
Table 9. Effects of design and order on participants’ choices. ..................................................... 92
Table 10. Summary of building design scenarios used in the evacuation simulation. ............... 110
Table 11. Performance of the examined models. ........................................................................ 111
Table 12. Estimation of the mixed logit model’s parameters. .................................................... 112
Table 13. Comparison of the neural network and mixed logit models in agents’ exit choices. . 114

ix

List of Figures
Fig. 1. Comparisons of the real and virtual metro station (left: real metro station, right: virtual
metro station). ............................................................................................................................... 30
Fig. 2. Illustration of the metro station layout, decision points and evacuation routes. ............... 30
Fig. 3. Illustration of the virtual fire and smoke. .......................................................................... 32
Fig. 4. Design strategy taken to manipulate visual access at DP 1. .............................................. 33
Fig. 5. Design strategy taken to manipulate visual access at DP 2. .............................................. 34
Fig. 6. Design strategy taken to manipulate visual access at DP 3. .............................................. 34
Fig. 7. Route choices of participants in Groups A and B and NPCs. ........................................... 41
Fig. 8. Directional choices of London, Beijing, and LA participants at DP 1 (* denotes the
direction with improved visual access). ........................................................................................ 42
Fig. 9. Directional choices of London, Beijing, and LA participants at DP 2 (* denotes the
direction with improved visual access). ........................................................................................ 43
Fig. 10. Route choices of participants in Groups C and D and NPCs. ......................................... 46
Fig. 11. Directional choices of participants at DP 1 in Groups C and D (* denotes the direction
with improved visual access in Group D, + denotes the direction that the majority of NPCs took).
....................................................................................................................................................... 46
Fig. 12. Directional choices of participants at DP 2 (+ denotes the direction that the majority of
NPCs took; * denotes the direction with improved visual access in Group D). ........................... 47
Fig. 13. Directional choices of participants at DP 3 in Groups C and D (* denotes the direction
with improved visual access in Group D, + denotes the direction that the majority of NPCs took).
....................................................................................................................................................... 48
Fig. 14. Participants’ evacuation trajectories. ............................................................................... 50
x

Fig. 15. The data analysis pipeline................................................................................................ 63
Fig. 16. Themes and subthemes of the focus group interviews. ................................................... 64
Fig. 17. Influence of countermeasures on occupant behavior during active shooter incidents. ... 73
Fig. 18. First-floor layouts of the standard office (left) and school (right). Stars denote
participants’ starting locations: one in the cafeteria and one in the hallway. Arrowed lines denote
the shooter’s movement trajectory. ............................................................................................... 80
Fig. 19. Virtual occupants and the virtual shooter. (a) occupants from the view of the starting
location in the cafeteria. (b) occupants from the view of the starting location in the hallway. (c)
shooter entering the building from the main entrance. ................................................................. 82
Fig. 20. Virtual tour guide for the building tour. .......................................................................... 86
Fig. 21. The effect of valence × time interaction on participants’ emotional arousals. ................ 88
Fig. 22. The effects of building × order × occupation interaction and design × order interaction
on participants’ choices................................................................................................................. 91
Fig. 23. The effect of building × occupation interaction on participants’ perceptions of other
occupants....................................................................................................................................... 93
Fig. 24. The effects of building × order × occupation interaction and design × occupation
interaction on participants’ ratings of occupant influence. ........................................................... 95
Fig. 25. The effect of building × occupation interaction on participants’ consideration of hiding
places............................................................................................................................................. 96
Fig. 26. Proposed framework for the development of behavioral data-driven agent-based
evacuation simulation. ................................................................................................................ 103
Fig. 27. Different levels of visual access in the virtual metro station (a) low, (b) medium, (c) low,
(d) high. ....................................................................................................................................... 104
xi

Fig. 28. Building layouts used in the evacuation simulation (walls are shown in white lines,
staircases are shown in gray rectangles, and arrows denote the moving direction on staircases.
Agents can only access the areas bounded by outer walls). ....................................................... 110
Fig. 29. Exit choices of agents using the neural network and mixed logit models for each
building design scenario. ............................................................................................................ 114
Fig. 30. Exit choices of agents using the HS model for each building layout. ........................... 116
Fig. 31. Evacuation time of agents using the neural network and mixed logit models for each
building design scenario. ............................................................................................................ 117
Fig. 32. Evacuation time of agents using the HS model for each building layout...................... 118

xii

Extended Abstract
Various emergencies of either natural or man-made origins could occur in built
environments, causing serious damage and putting human safety at great peril. For example, in
2018 alone, there were 379,600 residential building fires in the United States, resulting in 2,790
civilian deaths, 11,525 injuries, and $8,194,500,000 economic loss [1]. Moreover, a total of 277
active shooter incidents occurred in the United States between 2000 and 2018, which led to 884
deaths and 1,546 injuries [2]. Among these active shooter incidents, only 14% took place in open
spaces, while the remaining incidents happened in various built environments, including
commercial, educational, and government buildings. As people spend the majority of their time in
indoor environments [3], human safety in built environments, especially during emergencies, has
become an increasingly important issue. Available time and capacity of evacuation routes could
both be limited when an emergency occurs, hence effective response is extremely important [4–
6]. Among a variety of influencing aspects for effective emergency response, human factors are
critical [7]. It has been shown that in many emergency situations, non-adaptive human behavior
(e.g., pushing and stampeding) resulted in many undesired consequences. In fact, people interact
with people in different roles (e.g., other occupants, first responder teams, adversaries), building
attributes (e.g., location and visibility of signage, stairs, exits, etc.), and hazards (e.g., fires, smoke,
explosions) [8] during emergencies in built environments. Hence, having a comprehensive
understanding of human-building-emergency interactions is critical to mitigate the negative impact
of emergencies.
When an emergency occurs, it is crucial for people to quickly make decisions and respond
appropriately, such as “drop, cover, and hold on” during earthquakes, “run, hide, fight” during
active shooter incidents, and evacuating to a safe place during fires. For people to be able to
xiii

perform such actions, wayfinding, denoting “man's ability to reach spatial destinations in novel as
well as in familiar settings” [9], is of great importance [8]. Many social and environmental factors
can impact wayfinding behavior. Examples include signage and corridor configuration, crowd
flow, and people’s familiarity with the built environment [10–13]. Past research also evidenced
that participants are likely to exhibit distinct wayfinding behavior under different visual access
conditions [14], which are contingent upon a variety of factors, such as presence of hazards (e.g.,
smoke), indoor lighting conditions, individual’s location and visual acuity, as well as building
design [15–17]. Nevertheless, there still exists a gap about how different levels of architectural
visual access (i.e., visual access influenced by building designs) affect people’s wayfinding
behavior, including directional and route choices as well as people’s tendencies of following others.
As stated above, emergency scenarios are important determinants of human-building-
emergency interactions. Through the decades-long research efforts, most prior studies focused on
fires, whereas other types of emergencies have not widely come under scrutiny [5]. As active
shooter incidents present an increasing threat to the American society, especially in commercial
and educational environments [2], it is imperative to look into how people behave during these
incidents. Compared with other types of emergencies, active shooter incidents have several distinct
characteristics. First, human adversaries are present in active shooter incidents, who usually have
specified targets (e.g., building occupants) and can strategically respond to the ongoing situation.
On contrary, natural hazards (e.g., fires) are usually target-neutral and follow certain intrinsic
propagation rules. Second, while fires in built environments can last for hours [18], the duration
of active shooter incidents is usually short [19]. Third, crowd behavior during active shooter
incidents could be more complex and dynamic compared with other emergencies. Therefore,
having a deeper understanding of the preparedness of built environments in response to active
xiv

shooter incidents, as well as how the implementation of countermeasures affect human behavior
is of critical importance.
The ultimate goal of efforts in this research domain, regardless of the nature of emergencies
and the context of built environments, is to improve human safety during building emergencies
[20]. To design safer built environments and improve the preparedness for emergencies, the
behavioral design approach, which relies on the empirical knowledge about human-building-
emergency interactions, has been proposed [21]. Nevertheless, current tools for simulating
people’s evacuation behavior and building safety design are often based on over-simplified
assumptions of human behavior and ignore complex interactions among humans, buildings, and
emergencies. With the support of empirical human behavioral data during building emergencies,
there is an opportunity to identify and quantify such interactions.
Driven by the above-mentioned motivations, the research objectives of this dissertation are
defined as follows: (1) To understand the influence of architectural visual access on wayfinding
behavior in building fires, (2) To understand how different countermeasures influence building
security and human safety in response to active shooter incidents, (3) To empirically assess the
influence of countermeasures on human behavior during active shooter incidents in office and
school buildings, and (4) To leverage empirical knowledge about human-building-emergency
interactions and develop data-driven crowd evacuation simulations for informing building safety
design.
To address research objective 1, Chapter 4 presents the first study of this dissertation. A
virtual metro station based on a real station in Beijing, China was created as the experimental
environment. Multiple decision points (e.g., where participants needed to make directional choices)
were included in the virtual station. To manipulate the levels of architectural visual access,
xv

different design strategies (e.g., changing wall materials, removing columns in the hallway,
relocating ticket booths, etc.) were employed, which resulted in a different version of the virtual
station with varying conditions of architectural visual access at each decision point. To represent
crowd flow in the virtual station, non-player characters (NPCs) were included with
preprogrammed evacuation route choices. Moreover, to investigate potential influence of
participants’ cultural backgrounds on their wayfinding behavior, we conducted the experiments in
three different locations: London, United Kingdom, Beijing, China, and Los Angeles, United
States. A total of 226 participants were recruited. The results indicated that improving architectural
visual access could improve participants’ evacuation performance in building fires. It could also
influence participants’ directional choices, depending on the design strategy used and the spatial
characteristics of the building. In addition, participants’ tendency of following the crowd was
reduced when there was an alternative route with high architectural visual access. Furthermore,
the results suggested that whether cultural background could impact people’s wayfinding behavior
in building fires should be further investigated.
The second study of this dissertation, in which focus group interviews were conducted to
understand the preparedness of built environments in response to active shooter incidents, is
described in Chapter 5 to address research objective 2. The main goal of this study is to assess
current security countermeasures and identify various considerations associated with the
implementation of countermeasures. Fifteen participants with expertise and experience in diverse
operational and organizational backgrounds, including security, engineering, law enforcement,
emergency management and policy making, participated in three focus group interviews. A list of
countermeasures (e.g., access control, mass notification, etc.) that have been employed in response
to active shooter incidents was identified. Moreover, results from the focus group interviews
xvi

revealed important factors for successful implementation of countermeasures, including the
influence on human behavior, as well as occupants’ and administrators’ awareness of how to
appropriately use the countermeasures. The nature of incidents (e.g., internal vs. external threats),
types of built environments (e.g., office buildings vs. school buildings), and occupant
characteristics (e.g., students at different ages) were also recognized to influence the applicability
of countermeasures. The nexus between emergency preparedness and normal operations, and
tradeoffs among cost, aesthetics, and maintenance needs were also discussed. Additionally, to
ensure the effectiveness of countermeasures and improve human safety, the importance of training
for occupants, administrators, as well as first responder teams was highlighted.
The findings in Chapter 5 demonstrated the necessity and importance of research objective
3: Empirical assessment of countermeasures in how they affect human behavior during active
shooter incidents. Therefore, Chapter 6 aims to address the third objective. Office and school
buildings were selected as the experimental environments, as commercial and educational
environments have been most frequently targeted by shooters in the United States [2]. Based on a
comprehensive literature review and the findings in Chapter 5, a list of countermeasures in
response to active shooter incidents, including eliminating hiding places, installing barriers,
isolating unsecured areas from secured areas, access control, using frosted glasses, and staggering
interior doors, were identified and examined. Four different virtual environments, namely (1)
standard (i.e., without implementation of countermeasures) office, (2) enhanced (i.e., with
implementation of countermeasures) office, (3) standard school, and (4) enhanced school, were
created. A total of 162 office workers and middle/high school teachers were recruited to participate
in the virtual experiment. Each participant responded to the active shooter incident two times in
the office and school buildings with randomized order. The findings in Chapter 6 revealed that
xvii

implementing countermeasures for active shooter incidents could not only affect building security,
but also people’s decisions (e.g., run, hide, fight) and response performance (e.g., evacuation time)
when an active shooter incident occurs. It was also found in Chapter 6 that how people respond to
active shooter incidents are correlated with their daily roles as well as building and social contexts.
Teachers had more concerns of others’ safety compared with office workers. Moreover, teachers
had more positive perception of occupants in the school, whereas office workers had more positive
perception of occupants in the office.
Chapter 7 leverages the empirical findings in Chapter 4 and addresses the fourth objective
of developing data-driven crowd evacuation simulations for informing building safety design.
Social (i.e., crowd flow) and environmental (i.e., visual access and vertical movement) factors that
affected participants’ wayfinding behavior in Chapter 4 were extracted by reviewing the
experimental recordings. Different machine learning (e.g., k-nearest neighbors, support vector
machine, random forest, and neural network) and discrete choice models were employed to predict
the influence of social and environmental factors on people’s wayfinding behavior. The trained
models as well as a discrete choice model from a prior study were then used to develop agent-
based crowd evacuation simulations that incorporate agents’ strategic-level, tactical-level, and
operational-level decisions. The crowd evacuation simulation was further employed to examine
evacuation performance under different building design scenarios (i.e., different levels of visual
access, number and location of exits and staircases). The results showed that both the machine
learning and discrete choice models could accurately predict people’s directional choices during
emergency evacuations. Different building attributes could collectively influence human behavior,
leading to distinct exit choices and evacuation times. While both the trained machine learning and
discrete choice models generated similar results in evacuation simulations, the discrete choice
xviii

model had more interpretable results. By comparing the trained models in this study with a discrete
choice model developed in a prior study, it was found that agents had significantly distinct
responses to different building design options. Critical factors (e.g., type and size of buildings,
people’s familiarity with the building) for the applicability of evacuation models were identified.
Furthermore, recommendations were provided for future research that aims at employing crowd
evacuation simulations for building design evaluation and optimization.

1

Chapter 1. Introduction
Enhancing human safety during emergencies is a critical goal for both building design and
emergency management. Previous incidents have evidenced that inappropriate design choices
could result in undesired emergency consequences. For instance, during the Daegu Subway fire of
2003, the lack of emergency lighting and signage greatly increased the difficulty of efficient
evacuation [22]. Similarly, inappropriate responses during building emergencies could cause
disastrous outcomes. In the 2003 Rhode Island station nightclub fire, evacuees simultaneously
headed towards the main exit and ignored other available exits, which caused severe blockage and
resulted in dozens of fatalities [23]. These examples clearly demonstrated the interdependence
among humans, buildings and emergencies. When there are surrounding people during building
emergencies, interactions among people of the same or different roles (e.g., occupants, first
responder teams) occur frequently. For example, mutual aid and cooperation are common in
building emergencies, even among strangers [24]. Nevertheless, competing and selfish behavior
can also happen, due to increased stress and loss of personal space [25,26]. Moreover, humans and
buildings are two interwoven components. Building layout (e.g., location and number of rooms,
space adjacency, and location and number of exits) could affect people’s evacuation time, and
people’s knowledge of the building layout could largely influence their evacuation efficiency
[8,27]. Emergency attributes are also correlated with human behavior and building performance.
For example, in building fires, the presence of smoke imposes physiological impact and influences
people’s evacuation strategies and abilities of using spatial knowledge [28,29]. Moreover, human
behavior varies depending on the types of emergencies (e.g., fires, earthquakes, and active shooter
incidents) [30]. Therefore, it is important to have a comprehensive understanding of human-
building-emergency interactions to effectively improve human safety via building safety design
2

and emergency management.
Interaction defined as “reciprocal action or influence” of people and/or things on each other
[31], is the driving force of human behavior and safety during building emergencies, with human-
human interactions, human-building interactions, human-emergency interactions, and human-
building-emergency interactions contributing to its formation. In this dissertation, human-human
interactions refer to the collective behavior among building occupants and their interactions with
people in different roles, such as building administrators and first responder teams. Human-
building interactions refer to how various building attributes (e.g., signage, exits, stairs) impact
human behavior and how human behavior (e.g., using familiar exits, choosing stairs or elevators)
affects building performance during emergencies. Human-emergency interactions refer to how
emergency scenarios (e.g., presence of fire and smoke) impact human behavior and how people
cope with emergencies (e.g., extinguishing fire, fight with adversaries). Human-building-
emergency interactions refer to second-order interactions among humans, buildings, and
emergencies.
The study of human-building-emergency interactions dated back to the 1950s. At the initial
stage of this research domain, the driving force for human behavior during building emergencies
was considered to be panic, whereas the influences of social and environmental factors were not
taken into consideration [32]. Nevertheless, researchers later found that pure irrational behavior
resulted from panic is rather rare during emergencies [33]. Hence, following studies looked into
human-building-emergency interactions from a holistic view instead of considering them
irrelevant. Specifically, various theories and findings on human-human interactions have been
developed. For instance, past studies found that people are influenced by their social affiliations
and exhibit grouping behavior during building emergencies [34]. People were found to consider
3

themselves as members of social groups who are all threatened by emergency situations, behaving
on the basis of their social identity and act collectively [35]. Similarly, studies have evidenced that
people have the tendency to act in certain ways simply because others do during emergencies, as
a result of social influence and the intend to seek social proof [36]. Although these studies
investigated a wide spectrum of human-human interactions and developed theoretical foundations,
how human-human interactions would affect people’s response performance during building
emergencies are not well understood yet. For instance, while it was suggested that cooperation
could facilitate the evacuation process, another study that used evacuation drills showed that
behaving cooperatively would lengthen the evacuation time, since participants were too careful
not to push each other [37]. Such example indicates a main characteristic of this research domain,
namely the context dependency of research findings. Under different building emergency
scenarios (e.g., occupant density, familiarity with the building, etc.), the findings on human-
building-emergency interactions could be rather different. Therefore, collecting empirical data
about human behavior during building emergencies and conducting cross-validations are
important for discovering new findings and apply existing findings to different scenarios.
With regards to human-building interactions, past studies mostly targeted at the influence
of various building attributes on people’s wayfinding behavior during building emergencies.
Commonly examined building attributes include the configuration, color, and location of signage,
types of corridors, number and size of exits and so on [10,11,38]. However, other building
attributes, such as visual access and building layout, could also affect human behavior but were
insufficiently investigated in the literature. Possible reasons are twofold. First, many prior studies
simplified human-building interactions by assuming people would behave the same in different
types of built environments. Second, the study of human-building interactions often requires
4

manipulation of built environments, whereas many research methods, such as emergency drills,
post-emergency interviews, laboratory experiments lack such capabilities. Therefore, adopting
effective research methods that support investigating how people respond to emergencies in
different built environments are critical to advance our knowledge in this research domain.
As for human-emergency interactions, unspecified emergencies and fires were examined
by the majority of prior studies. How people interact with other types of emergencies that can
occur in built environments were much less investigated. Specifically, there have been very few
empirical findings related to human behavior during active shooter incidents. Although more
research in this direction has been developed with the increasing number of active shooter
incidents in the U.S. [2], the research focus has been lying in designing anti-terrorism buildings,
developing response procedures, and educating people the recommended behavior [39–41].
Furthermore, the large number of prior studies that focused on unspecified emergencies indicates
that these studies assume similarities of human behavior during building emergencies. In fact,
human-emergency interactions may vary in different situations due to different emergency
characteristics and response goals. For example, during active shooter incidents, the recommended
responses are “run, hide, fight”; whereas in earthquakes, people are suggested to follow the “drop,
cover, hold on” procedure.
In addition to the three types of interactions discussed above, only a few studies
investigated interactions between people and certain building attributes, such as corridors and
signage [10], elevators and stairs [42], and exits [43] in normal and emergency situations. As a
result, humans, buildings, and emergencies should be considered collectively to have a holistic
understanding of their relationships and leverage the knowledge to improve human safety during
building emergencies.
5

During the past few decades, there is a growing trend in the usage of crowd evacuation
simulations for building safety design [44]. Various evacuation simulation tools have been
developed and adopted in the safety design of different building types [45]. Accurate prediction of
human behavior during building emergencies is of vital importance for applying crowd evacuation
simulations. Nevertheless, many prior studies only considered how basic attributes (e.g., distance
to exits, number of exits, etc.) of surrounding environments could affect human behavior
[27,46,47]. Common approaches for crowd evacuation simulations are to define rules based on
extant behavioral theories and to refine established models [26,48]. As behavioral data is
considered crucial, incorporating empirical findings about human-building-emergency
interactions could enhance the reliability and validity of crowd evacuation simulations.
The overarching goal of this dissertation is to deepen our understanding of human-
building-emergency interactions that involve people from various backgrounds, different types of
built environments and emergencies, as well as to leverage the empirical findings to develop data-
driven crowd evacuation simulations for informing building safety design. The structure of this
dissertation is organized as follows: Chapter 1 provides the overall background and motivation for
the research efforts. A thorough literature review related to the scope of this dissertation is
described in Chapter 2. The research objectives and questions are presented in Chapter 3. Chapter
4 presents the first study of this dissertation that investigates the influence of architectural visual
access on people’s wayfinding behavior in building fires. The studies that focus on implementing
countermeasures in different built environments in response to active shooter incidents, and how
countermeasures influence human behavior during active shooter incidents are illustrated in
Chapter 5 and Chapter 6, respectively. Chapter 7 presents the study that leverages empirical
findings about human-building-emergency interactions to develop crowd evacuation simulations
6

for informing building safety design. Finally, Chapter 8 concludes the dissertation by summarizing
the main findings and discussing the limitations and future directions.

7

Chapter 2. Literature Review
This chapter thoroughly reviews prior studies related to this dissertation. Various types of
human-building-emergency interactions are presented in sections 2.1 – 2.3. Sections 2.4 provides
a literature review for the influence of visual access on emergency wayfinding (Chapter 4). Section
2.5 reviews the studies on the implementation of security countermeasures in response to active
shooter incidents along with the influence on human behavior, which is related to the studies
described in Chapter 5 and Chapter 6. Section 2.6 presents a review of the literature that focused
on crowd evacuation simulations and Section 2.7 reviews the application of crowd evacuation
simulations for building safety design.
2.1 Human-human interactions
In many circumstances, people are not alone during building emergencies. They may be
accompanied by others, such as their families, coworkers, or strangers. Human-human interactions
are one of the most important aspects that determine how people behave and the overall evacuation
time and patterns [49].
2.1.1 Occupant-occupant interactions
Human-human interactions are crucial determinants of human behavior during building
emergencies. Especially in large public buildings, people tend to observe others’ responses and
behave accordingly [50]. Typical types of human-human interactions include herding, avoiding,
grouping, helping and competing, leader-following, and information sharing [5,26,51,52]. Herding
behavior is a type of interactive behavior. It refers to a person following what others are doing,
even though the perceived situational information suggests otherwise [53]. With respect to
emergency evacuation, herding behavior refers to an evacuee choosing the most congested route
because that route is the most popular choice, instead of alternative routes with less people [13].
8

In the early 2000s, it was suggested that herding behavior would occur when people experience
high levels of stress [54]. However, subsequent studies revealed that herding behavior could be a
result of rational decision-making process and it is related to the lack of information that people
need to understand the situation and make a decision [13]. Recent studies further evidenced that
herding behavior is impacted by both environmental factors (e.g., number of evacuees near exits,
exit visibility, crowd density) and personal factors (e.g., herding attitudes) [13,55,56]. With regard
to its influence, on the one hand, herding behavior may facilitate the evacuation of those who are
not familiar with the building, representing a type of cooperation in which people share their
knowledge [57]. On the other hand, it may also lead to inefficient exit choice and decrease
evacuation efficiency [26,58]. In addition to following the crowd, prior research also pointed out
that people might prefer to have personal space and avoid physical contact with others during
emergencies [26]. Avoiding behavior is also related to building attributes and environmental
factors. Studies that conducted laboratory experiments illustrated that in highly crowded places
with low uncertainty (e.g., no obstacles blocking visibility), people would avoid choosing the same
direction as the majority [55]. Moreover, when exits with shorter distance were overcrowded, the
majority of people would tend to choose further exits to avoid excessive delays due to heavy
congestions [59].
Similar to the herding and avoiding behavior, grouping behavior is another type of
interactive behavior that involves multiple people. While herding and avoiding behavior may occur
among crowds of strangers, grouping behavior is usually based on some form of social
connectedness [35]. When people are with families, close co-workers or friends, they tend to move
as a group and even re-enter the building to search for missing members [60,61]. Grouping
behavior frequently occur in real-world emergencies, including the Beverly Hills Supper Club Fire
9

in 1977 [62], which demonstrated that the group size had an influence on evacuation efficiency.
Additionally, grouping behavior has been incorporated in simulations to analyze its impact on the
evacuation process. It was found that the evacuation time would be significantly prolonged if
evacuees travelled back and forth and took detours to seek group members [49].
Compared with the grouping behavior, helping and competing behavior are also related to
people’s pre-existing social, as well as their emergent collective identities during building
emergencies [62]. Contrary to the panic theory, people often exhibit helping behavior during
building emergencies, which is observed in many real-world building emergencies, such as the
July 7
th
London bombings in 2005 [63] and the Rhode Island station nightclub fire in 2003 [23].
Collective bonds among people might be strengthened and even created through the experience of
an emergency, and higher collective identification increases cooperation among people, while
higher level of danger decreases the amount of help [64]. On contrary, competing and selfish
behavior can also happen, due to increased stress and loss of personal space [25,26]. The presence
of competing behavior could result in more physical collisions, clogged exits, and inefficient
evacuations, as reported in several studies that conducted simulated evacuations [26,65].
Additionally, based on the survey data from 1,134 respondents in a train station in Melbourne,
summarized in a study published in 2017, it was shown that men were more likely to behave
competitively than women [66]. However, the case study of the Beverly Hills Supper Club Fire
demonstrated that men helped women more often than women helped men [67], which may
suggest a change in human behavior over time.
Apart from pre-existing or emergent social relationships, people’s behavior is influenced
by their social roles in their daily lives as well [68]. Thus, people can take roles of leaders and
followers when emergencies happen based on their personality, knowledge and experience, and
10

social roles in their daily lives. Most people adopt the role of followers during emergencies and
respond after others’ actions [8]. Leaders may be authority figures, individuals and social groups,
and they can lead followers to perceive environmental cues as well as to guide their evacuation
process [5,69]. Moreover, queuing behavior can be observed when a leader slows down or stops
and followers form a waiting line [70]. Leader-following and helping behavior can also have
coupled effects: when altruistic leaders slow down to help injuries, followers reduce their speed
accordingly, which slows down the evacuation process [71].
Situational information plays an important role in emergencies and sometimes acts as a
medium in human-human interactions. During building emergencies, people are “information
hungry” and make efforts to gain more information about the emergency situation, such as
consulting others and forming a group to discuss the situation [72]. A survey targeting at human
behavior in fires in high-rise residential buildings also revealed that most responders would warn
others and/or ask neighbors if there was a fire [73]. Information sharing acts as a determinant of
people’s evaluation of the situation, intention to act, and evacuation route choices. It enables
people to share situational information during emergencies (e.g., infeasible evacuation path), so
that they can take more appropriate actions accordingly. Nevertheless, information sharing could
also prolong the pre-movement time (i.e., the delay time from the perception of emergency cues
to the movement to a safe place, typically to an exit) and delay the evacuation process [69].
2.1.2 Occupant-staff interactions
Staff, such as security personnel and first responders, are often present in building
emergencies and play important roles [74]. Sime, in his study of affiliative behavior [34],
highlighted the difference between staff behavior and occupant behavior. During emergency
evacuations, staff and building occupants tend to follow different egress routes due to different
11

levels of familiarity with the building [75]. Therefore, occupant-staff interactions have different
characteristics compared with the interactions among occupants. It was demonstrated that many
people tend to seek information or wait for directions from staff members before taking actions
during emergencies [66]. Occupants with disabilities particularly trust and rely on staff members
when an emergency occurs [76]. It was also noted that staff, who have more emergency drills and
training experiences, could respond properly and inform building occupants to evacuate
immediately [77,78]. Moreover, staff could engage in many alternative activities that facilitate
evacuation, such as directing occupants to exits, giving out supplies (e.g., water), and helping the
injured [30,69]. In the Rhode Island station nightclub fire, seven of the twelve staff, including
bartenders, bouncers and waitresses, were involved into helping the occupants [23]. Nevertheless,
while staff can provide help to occupants and guide them to behave appropriately, it has been noted
that counter-flow may also be caused by the movement of staff and building occupants [69]. For
example, fire fighters and rescuers who run into the buildings and move against evacuees may
cause evacuation delays [74,79].
2.2 Human-building interactions
Buildings provide primary conditions for the possibility of surviving an emergency [8].
There exist complex interactions between humans and various building attributes, which could
impact human safety and a building’s performance during emergencies.
2.2.1 Human-signage interactions
Signage systems have long been regarded as one of the most important building attributes
in both normal conditions and during emergencies [11,80]. However, the effectiveness of signage
systems does not always reach expectations. A study demonstrated that only 38% of people
perceived signage systems and used the information for evacuation [81]. Thus, how to improve
12

the effectiveness of signage during building emergencies have been widely studied, such as
changing color and location of signage, and using dissuasive exit signage [38,82,83]. It was found
that people rarely perceived signage installed at the ceiling level, while signage located at the floor
level were more effective, especially with dense smoke [84]. Additionally, the interactions
between people and signage systems are related to cognitive factors (e.g., interpretation of the
information conveyed by the signage) and psychological factors (e.g., desire to believe the
information) [81]. Cultural and local implications are thus an important consideration, as there are
different signage configurations worldwide, which might cause people from different cultures and
locations interpret the meaning of signage differently. For instance, most international building
codes prescribe exit signage to be green, whereas certain codes also allow users to choose between
green and red exit signage [38]. Green and red exit signage were found to have similar connotations
for both Chinese and European participants during emergency evacuation [85]. Moreover,
compared with local exposure (e.g., signage color in the local environment where people reside),
semantic association (e.g., green = exit) was found more influential to people’s exit choice [38].
To further magnify the effectiveness of signage systems, researchers developed active dynamic
signage that could provide adaptive information (e.g., fire propagation), exclude unsafe routes and
guide people to safe places [86]. Galea et al. (2017) found that the active dynamic signage could
be correctly interpreted by most of the respondents from across the world and was able to direct
most people to a distant exit and keep them away from a closer but non-viable exit.
2.2.2 Human-exit interactions
Exits are one of the most fundamental building attributes, especially during emergencies.
Earlier research have set several basic engineering features for exits, such as maximum flow rate
capacity and required number of exits [8]. Thereafter, many studies adopted a more behavioral
13

perspective and examined how people interact with exits during emergencies. Some researchers
concluded that the unbalanced usage of exits during emergencies was related to exit locations [88].
If exits are open and people can see the outside, these exits are more attractive and likely to be
chosen more frequently [89]. Moreover, exit choice is related to an individual’s role. Compared
with staff members, building occupants mostly egress through the main building exits instead of
the emergency exits, due to their insufficient knowledge of the building and lack of prior
emergency training experience [34]. Moreover, People tend to evacuate the building using exits
that they are most familiar with (e.g., the main entrance of the building) since routes leading to
familiar exits are often perceived as the shortest [8,12]. That being said, another study that
conducted emergency evacuation experiments in a two-dimensional virtual environment found
that participants did not have any preference for their familiar exits [90]. Whether this result can
amount to occupant behavior in real-world emergencies, however, is debatable, as the participants
only had a top-down view of the virtual environment and no hazard was included. Furthermore,
the combined effect of exits and other building attributes was studied. It was suggested that to
facilitate the evacuation process, obstacles near exits should be cleared, gathering places should
not be close to exits and more signage is needed near exits [78].
2.2.3 Human-corridor interactions
Corridors are essential components for horizontal accessibility during building
emergencies. The flow rate of a corridor is a significant indicator of evacuation performance [8].
Many aspects of corridor configuration have been studied. For example, if a corridor contains a
widening, it could result in disturbances instead of increasing the flow rate, in that people would
increase the distance from each other and squeeze in at the end of the widening [54]. Empirical
data from human and ant experiments showed that with the presence of high-density crowd, higher
14

turning angles of corridors were inefficient because they reduced the flow rates during evacuations.
When stress level is high, the reduction of flow rates was more significant due to the stress-induced
competing behavior [91]. With regard to merging corridors, a laboratory-controlled evacuation
experiment found that a symmetric angle was more efficient than an asymmetric setup with equal
angles when the participants were merging to a third corridor from two different corridors [92].
VR-based studies also showed that people preferred to follow brighter and wider pathways, and
their egress route choice was subject to the intersection type of corridors (e.g., T-type) when a
signage system was absent [10,93]. In addition, corridors with more turns and unfamiliar routes
were perceived to be longer, which decreased the probability of the corridor being chosen as part
of the evacuation route [8]. Due to the grouping behavior (e.g., waiting to stay close to group
members), congestions may be built up at the intersection of corridors, which lengthens the overall
evacuation time [94].
2.2.4 Human-vertical accessibility interactions
Stairs, elevators, and escalators are common building attributes for vertical accessibility.
Movement rates on stairs have been studied extensively in relation to the configuration of stairs
(e.g., effective width, spacing on stairs, etc.) [95,96]. The investigations on the 2001 WTC attack
revealed that adequate lighting on stairs, marked handrails, and steps with reflective tapes
facilitated the evacuation process, whereas debris on stairs and locked stair doors were barriers for
efficient evacuation [97]. Meanwhile, human behavior can impact the efficiency of movement on
stairs as well. A slow person entering the stairs and counter-flow are likely to decrease the overall
movement speed [79,98]. Merging streams of evacuees in the floor-stair intersection are also a
significant issue during evacuations. It was suggested that landing doors should be connected
opposite the incoming stairs instead of adjacent to the incoming stairs to improve the efficiency of
15

merging streams [99]. Apart from the stairs, using elevators and escalators is an alternative choice,
especially for those with disabilities or medical conditions. With more high-rise buildings built in
the recent years, a combined use of elevators and stairs is deemed practical to improve evacuation
efficiency [100]. However, prior studies have found that people’s tendency of using elevators and
escalators was low, due to the belief that it is safer to use stairs during emergencies [42,66]. Thus,
instead of assuming that people will use elevators to evacuate during emergencies, human behavior
must be taken into consideration when designing building evacuation systems that utilize elevators.
Significant differences in elevator/stairs choice during emergencies were found between the U.S.
and Chinese respondents: 52.5% of the U.S. respondents considered using elevators during
emergencies, which was around twice the proportion of Chinese responders (21.5%) [42]. The
difference indicates there might be cultural impacts in terms of using elevators during emergencies,
thus further cross-cultural investigations are necessary.
2.2.5 Human-alarm interactions
Emergency alarms and announcements are important information sources during
emergencies, especially at the early stage when only ambiguous information is perceivable [50].
In the last century, emergency announcements were seldom used to provide emergency
information because of the false belief that people would behave irrationally if they knew there
was an emergency. However, it was revealed by more recent studies that telling the truth about the
emergency could in fact motivate people to start evacuating more quickly and shorten the pre-
movement time, which accounts for a large proportion of the total evacuation time [50].
Nevertheless, in real-world emergencies, accurate notifications might not always be present or
possible [101]. For example, during the 2001 WTC attack, a building-wide announcement in WTC
2 assured that the building was safe and asked people to return to their offices [69]. Moreover,
16

many people tend to ignore alarms instead of taking immediate actions, particularly when they are
not near the hazard [102], when they have past experience of false alarms or frequent drills [103],
when they need more situational information, and when they are committed to other tasks [104].
An evaluation of different alarm types showed that announcements providing timely instructions
are more effective than siren alarms [105]. Additional alarm modes need to be redesigned,
especially for people with disabilities, since the signal type may not be adequate for them and can
instead result in undesired effects [76].
2.2.6 Other human-building interactions
Beyond the specific building attributes, several building characteristics that influence
emergency evacuations have been identified. These are: (1) visual access, (2) degree of
architectural differentiation (i.e., unique building characteristics that can be used for
orientation/wayfinding purposes) and (3) plan configuration [80]. Changes of spatial accessibility
(i.e., caused by activated fire shutters) were also found to have negative impacts if people do not
have enough awareness of the change of spatial accessibility during emergencies [75]. In addition,
as the environments get more complex with more visual and aural noise, the more it becomes
difficult for people to identify emergency cues due to more amount of irrelevant environmental
stimuli [106]. An example is hospitals, where some patients have limited mobility, hence the
efficiency of evacuations is likely to be compromised as well [73].
2.3 Human-emergency interactions
Human behavior is dependent on the emergency scenario. While all building emergencies
share certain common characteristics (e.g., causing stressful situations), the results of human
behavior studies specific to one type of emergency may not always be directly applicable to other
emergencies [30]. It is necessary to study how people interact with different emergencies.
17

2.3.1 Human-fire interactions
Fire is a widely studied type of emergency in buildings. Several fire attributes, including
perceptual attributes (visual, audible, tangible features and smell), fire growth rate, heat, smoke
yield, and toxicity have been identified to impact people’s emergency response performance [8].
However, people often ignore ambiguous fire cues (e.g., fire alarms) and continue with their
activities instead of starting to evacuate [8]. Typical coping strategies with fire include:
extinguishing fire, taking shelter to avoid fire, and evacuation [107]. Depending on the severity of
fire, both fighting the fire and avoiding the fire have been observed during fire emergencies [108].
In addition, smoke is a critical attribute that is often present in fire emergencies. However, people
do not always perceive smoke as evidence of fire. They may interpret smoke as a normal
phenomenon, such as smoke coming from a restaurant kitchen [89]. Through an experiment in
smoke-filled corridors, it was reported that high smoke density could significantly reduce evacuees’
speed, as well as their thinking power [109]. Smoke also imposes influence on spatial visibility. If
the visibility range is limited, people tend to follow the crowds [15,110]. However, contrary to the
conception that people are reluctant to move through smoke, studies of major incidents showed
that people were actually willing to move through smoke when they believed that they were
heading towards safety [111]. In dramatic fire emergencies, there are three situations that people
do not have to move through smoke: (1) they are located below the fire floor in a high-rise building;
(2) they are remote from the fire site in a large horizontal structure; and (3) they start the evacuation
early [111]. When confronted with dense smoke, people may also redirect their paths in order to
avoid breathing difficulty, lack of visibility, and out of fear [28].
2.3.2 Human-earthquake emergency interactions
During earthquake emergencies, people’s common responses include freezing in place,
18

evacuating the building immediately, taking cover, and protecting others and property [112]. An
investigation of the 2012 Northern Italy Earthquakes, which occurred at night, showed that
escaping from home, moving to another room of the house, going down the stairs, and getting
dressed were the most frequently reported behaviors in the immediate aftermath of the earthquake
[113]. Moreover, the earthquake magnitude greatly influences the emergency situation, and human
behavior is influenced by the post-earthquake damaged environments and their level of earthquake
preparedness [114]. Several earthquake attributes have been identified to have direct interference
with human behavior, namely (1) buildings shaking, (2) ruins and “high building” influence (i.e.,
avoid areas surrounded by high buildings after evacuating buildings), and (3) presence of visible
damage [21]. Post-earthquake building modifications can provoke several human behaviors, such
as fear of buildings (i.e., running out of buildings and keeping a distance from them) and social
attachment (e.g., helping behavior) [21]. Impediments to egress routes, other physical damages
(e.g., fallen ceiling tiles), and injured individuals blocking the routes were found to delay
earthquake evacuation based on the results of a simulated evacuation [115].
2.3.3 Human-violence interactions
In the acts of extreme violence (mass shootings, terrorist attacks, etc.), various weapon
types may be used (e.g., bombs, firearms, incendiary). Adversaries also evolve their techniques,
such as using more advanced weapons and strategies [40]. The 2001 WTC attack, as an act of
extreme violence, has been extensively investigated around the world and generated many
behavioral data. However, since no adversaries were present after the planes hit the WTC, how
people interact with adversaries have not been studied. There are few studies that looked into
adversary behavior during attacks. However, the existing studies that simulated adversary behavior
tended to ignore the decision-making procedure of adversaries, and instead considered their
19

behavior as essentially random [116], or simply choosing the nearest person as the target [117]. In
fact, unlike other emergencies, adversaries (such as shooters in an attack) are essential attributes
in acts of extreme violence, thus closer examination of the adversaries is of critical importance.
Moreover, suggested responses to acts of extreme violence have not been well defined [118].
While some studies simulated people’s movement in the presence of adversaries (e.g., people try
to keep away from adversaries) [119,120], more studies concerning fine-grained interactions
between building occupants and adversaries are needed to understand the influence of interactions
on human safety.
2.4 Influence of visual access on emergency wayfinding
Visual access in buildings could be affected by a variety of factors. To represent reduced
or zero visual access caused by smoke, eye-patches and glasses were commonly used in prior
studies to conduct human-subject experiments [110,121]. For example, Guo et al. [122] conducted
experiments in a classroom, where participants were asked to evacuate under conditions of zero or
good visual access by wearing or removing eye-patches. The results revealed that participants
tended to select routes unoccupied by others under good visual access, whereas they explored their
surroundings using their hands and bodies without considering the congestion level in the zero-
visual access condition. Failure of electricity supply systems could also greatly diminish visual
access during building emergencies. Jeon et al. [17] analyzed the evacuation process in an
underground transportation facility under different conditions of indoor lighting and smoke levels,
using eye-patches with different opaqueness. Their experiments showed that smoke had stronger
influence on participants’ evacuation performance (e.g., speed and distance) than indoor lighting.
Where people are located in a building is another influencing factor on the level of visual access
they could obtain. It has been shown that people tend to choose the exits that are visible to them
20

and that are open and they can see through [89].
Visual access can be influenced by building design as well (i.e., architectural visual access),
hence influencing human-building interactions. Gärling et al. [123] conducted evacuation
experiments in a university building and found that with lower level of architectural visual access
(openness of building layout), participants’ performance in the orientation test improved less over
time. Seidel [124] found that in the airport environment, wayfinding was easier for participants
arriving at the gate, if they had direct visual access to the baggage claim area. Different levels of
architectural visual access could also affect people’s route choices during their wayfinding process.
For example, Carpman et al. [125] conducted a video simulation experiment in a hospital
environment and found that compared with available signage, architectural visual access (presence
of an entrance) had a stronger influence on the participants’ route choices. In another study, it was
found that fire exit doors that were faced with murals, even though visible to the participants, might
not be perceived as doors and caused confusion for the participants [126]. That being said, how
different levels of architectural visual access affect people’s wayfinding behavior during
emergencies still remains underexplored [127]. Moreover, since various design strategies (e.g.,
locations of walls, columns, stairs) could affect visual access, multiple decision points with varying
architectural visual access conditions could be used to examine the influence of architectural visual
access on wayfinding behavior.
2.5 Building countermeasures in response to active shooter incidents
To mitigate the risks that active shooter incidents impose on buildings and occupants,
several public agencies have published guidelines and recommendations on preparedness,
response, and management related to active shooter incidents. For example, the Interagency
Security Committee released a document for planning and responding to an active shooter incident
21

and highlighted the importance of preparedness (e.g., establishing threat assessment teams) and
training (e.g., conducting drills) [128]. The FBI recommended that while there is no absolute best
response strategy during active shooter incidents, maintaining a “run, hide, fight” mindset can
increase occupant safety [41]. Another active shooter/hostile event guide compiled by the
Interagency Board underlined that an Incident Command System (ICS) is important to foster
incident management and coordination when multiple agencies are involved and suggested a
bottom-up approach to build ICS [129]. Research efforts have also been made to improve building
preparedness and investigate the relationship between buildings and occupants during active
shooter incidents. Kuligowski presented guidance on the creation and dissemination of emergency
information in both audible and visual means in response to active shooter incidents [130]. Based
on the results of agent-based simulations, Cho et al. showed that with the presence of human-
sensing technology and building information, the efficiency of safe evacuation can be significantly
improved during active shooter incidents [131]. Lee et al. demonstrated that alert systems in public
buildings, as well as quick responses of occupants and first responders are helpful to decrease the
casualty of active shooter incidents [132]. Public address systems and mass notifications (e.g., text
and email messages) were suggested to be deployed in classrooms, dormitories and outdoor
environments, in order to tackle active shooter incidents in university campuses [133,134]. Fox
and Savage also stressed that the effectiveness of countermeasures may be different on university
campuses as compared with high schools [135].
Apart from the above-mentioned efforts, several documents have been developed for the
purpose of guiding building design in preparation for active shooter incidents and other types of
attacks (e.g., explosive blasts, chemical, biological, and radiological attacks). These documents
include Primer to Design Safe School Projects in Case of Terrorist Attacks [136], Primer for
22

Design of Commercial Buildings to Mitigate Terrorist Attacks [137], Reference Manual to
Mitigate Potential Terrorist Attacks Against Buildings [118], and Minimum Antiterrorism
Standards for Buildings [138]. In these guidelines, lists of countermeasures have been proposed
for protecting buildings and occupants against various attacks. Nevertheless, these
countermeasures are frequently aimed at deterring attacks and mitigating the effect of attacks on
buildings and occupants, whereas when an attack (e.g., active shooter incident) occurs, the
effectiveness of these countermeasures, particularly how they influence the actions of occupants
and administrators, is yet to be investigated.
2.6 Crowd evacuation simulations
Crowd evacuation simulations provide an efficient approach to represent human behavior
during building emergencies and allow designers and engineers to quantitatively assess the
evacuation performance of different building design options. Santos and Aguirre [139] presented
a critical review about evacuation simulations and categorized them into flow-based models,
cellular automata models, and agent-based models. In a later review by Zheng et al. [51], seven
methodological approaches for evacuation simulations were identified: fluid-dynamics models,
cellular automata models, lattice gas models, social force models, agent-based models, game
theoretic models, and approaches based on experiments with animals. More recently, Chen et al.
[44] presented an up-to-date literature review and summarized four types of evacuation
simulations, namely macroscopic approach-based models, cellular automata models, social force
models, and agent-based models.
In macroscopic approach-based models, environments are generally represented as
networks (i.e., nodes connected by edges) and occupants are considered to behave homogeneously
in analogy to fluids and gas or governed by certain rules (e.g., moving to the destination with the
23

minimum estimated time) [44]. While macroscopic approach-based models can evaluate high-
level evacuation performance (e.g., flow rate), the main limitation is the lack of flexibility to
represent more fine-grained interactions among occupants [51]. Given this reason, despite that
macroscopic approach-based models were employed in early-stage evacuation simulation tools
(e.g., EXITT [140]), more contemporary evacuation simulation tools have turned to other models
for better simulation accuracy. Cellular automata models represent the environment as a regular
grid of cells. The state of cells evolves at a certain time interval based on the state of neighboring
cells and a set of local rules, such as distance to exits and crowd density [141,142]. Due to the high
computational efficiency, cellular automata models have been commonly used in evacuation
simulation tools, such as EXODUS [28] and STEPS [143]. Nevertheless, cellular automata models
have limited ability in representing heterogenous occupants [44]. Social force model was first
proposed by Helbing and Molnár and has been extensively adopted in evacuation simulations [144].
In social force model, three socio-psychological and physical forces, namely driving force to the
desired destination, interaction force with other occupants, and interaction force with walls, are
included and jointly determine the evacuation behavior. Since the structure of social force model
is relatively flexible, many researchers have introduced other influencing factors, such as the
leader-following behavior [48] and grouping behavior [145] in the original model. Evacuation
simulation tools that are based on social force models include FDS + Evac [146], PTV Viswalk
[147], and Simwalk PRO [148].
Agent-based models utilize a decentralized bottom-up approach to capture individual-level
behavior and emergent phenomena. Each individual is considered as an autonomous decision-
making entity with the ability to assess the current situation and make decisions based on certain
behavioral rules [149]. For instance, based on multiple social theories, Pan et al. [150] considered
24

various social behavior (e.g., following, queuing, competing) during evacuation in their agent-
based model. Similarly, von Sivers et al. [63] used insights entailed in social identity theory and
self-categorization theory, and incorporated helping behavior during evacuation in their agent-
based model. There is an increasing trend of agent-based models being employed in commercially
available evacuation simulation tools in recent years, including Pathfinder [151], MassMotion
[152], and Anylogic [153]. Due to the bottom-up structure of agent-based models, investigating
how individuals respond to building emergencies is crucial for improving the fidelity and
reliability of evacuation simulations. Therefore, many prior studies leveraged empirical data to
understand how individuals evaluate various factors when making decisions. For example, Busogi
et al. [154] conducted a survey to estimate occupants’ cost and reward estimates of different
evacuation means (e.g., using elevator, stairs, jumping, and climbing) and employed the Markov
Decision Process to represent agents’ decision-making process in their evacuation simulations.
Rozo et al. [155] used a discrete choice model to capture agents’ decision-making mechanism and
conducted emergency drills to calibrate the impact of distance to exits and crowd density on exit
choices. Similarly, Haghani and Sarvi [156] leveraged the data from laboratory experiments to
calibrate a discrete choice model that describes the influence of multiple social and environmental
factors on agents’ adaptive exit-choice changes.
So far, the development of agent-based models has been primarily grounded on pre-
determined rules, observations, and data from surveys, laboratory experiments, and emergency
drills. VR-based experiments were found to be an effective method for studying occupant behavior
during building emergencies and have been increasingly used in recent years [5]. However, few
studies leveraged findings in VR-based experiments to develop evacuation simulations and
evaluate building design options. In addition, prior studies applied discrete choice models to
25

describe agents’ decision-making process. While several recent studies leveraged machine
learning models for predicting occupants’ pre-evacuation actions and movement patterns during
evacuations [157,158], machine learning and discrete choice models were only compared based
on existing datasets. As the main goal of employing evacuation simulations is for building design
evaluation, machine learning and discrete choice models should be further examined in evacuation
simulations with different building design scenarios.
2.7 Crowd evacuation simulations for building safety design
Crowd evacuation simulations have been widely used for building safety design due to
their capability of producing quantitative results (e.g., number of casualties, evacuation time, exit
choices, and congestion maps) that facilitate the evaluation of different building design options.
Sun and Turkan [159] developed a BIM-based evacuation simulation for fire safety management
and validated the results by comparing with a real fire incident (i.e., the Station Nightclub fire in
2003). Fang et al. [160] conducted an experiment to examine occupants’ exit choices and
developed an evacuation simulation to analyze the usage of exits with different crowd density
levels. Similarly, Li et al. [161] leveraged the congestion map outputs of evacuation simulations
to analyze the spatial and temporal distribution of congestion risks on escalators. In prior studies,
building attributes most often examined included the number, location, and size of doors and exits,
hallway width, and presence of obstacles. For instance, Ha and Lykotrafitis [27] developed an
evacuation simulation for a multi-room multi-floor building and investigated the impact of the
width of doors and exits on evacuation performance. Their results suggested that the size of main
exits is the most important attribute that determines evacuation times. Arteaga and Park [46]
examined the impact of door, exit, and hallway widths on the number of casualties during an active
shooter incident in their evacuation simulation. They observed that hallway and door widths could
26

bring the largest improvement for occupant safety. Furthermore, Hu et al. [47] proposed a machine
learning framework to learn the relationship between occupants’ evacuation performance and
different building attributes (e.g., exit location, presence of obstacles, etc.) through evacuation
simulations.
Despite that these attributes (e.g., doors, exits, hallways) are essential for a building and
have critical impacts during emergency evacuations, prior studies lack the consideration of high-
level building attributes (e.g., visual access) and complex building layouts (e.g., multiple floors)
when employing evacuation simulations for building safety design. Hence, the evacuation process
during building emergencies is often simplified to provide a baseline for evacuation performance
[162]. Additionally, the validity of utilizing evacuation simulations for building safety design has
been insufficiently considered. Many studies presented the effect of different building attributes
without providing validation for the results [27,46]. For those studies that considered validation,
commonly used approaches include macroscopic examination using the fundamental diagram (i.e.,
relationship between flow and density) and comparison with past emergency incidents [48,159].
Since the data about real-world incidents may not be readily accessible, comparing with evidence
collected in human-subjects experiments is an alternative approach for validation purposes [44],
which was adopted in this study. Moreover, as presented in section 2.1, occupant behavior during
building emergencies is contingent upon a variety of factors. The findings about how building
attributes affect evacuation performance may not be directly applied in other contexts. That being
said, most prior studies drew conclusions about building safety design using a single evacuation
simulation without presenting specifics of the emergency scenario or implications about
employing the evacuation simulation in other contexts.

27

Chapter 3. Research Objectives and Questions
3.1 Research objective 1
To understand the influence of architectural visual access on wayfinding behavior in
building fires.
• Research Question 1.1: How do different levels of architectural visual access impact
people’s route and directional choices during the evacuation process in building fires?
• Research Question 1.2: How do different levels of architectural visual access impact
people’s evacuation performance in building fires?
• Research Question 1.3: What is the interaction impact of architectural visual access and
crowd flow on wayfinding behavior in building fires?
• Research Question 1.4: How do different cultural backgrounds affect wayfinding behavior
under different level of architectural visual access in building fires?
3.2 Research objective 2
To understand how different countermeasures influence building security and occupant
safety in response to active shooter incidents.
• Research Question 2.1: What are the countermeasures that have been used to date to proof
buildings against active shooter incidents?
• Research Question 2.2: How do different countermeasures influence the behavior of
occupants, shooters, and first responder teams during active shooter incidents?
• Research Question 2.3: How do different contextual factors (e.g., internal vs external
threats, building and occupant types) affect countermeasure implementations to proof
buildings against active shooter incidents?
• Research Question 2.4: What are the trade-offs among countermeasure implementations in
28

response to active shooter incidents, other types of emergencies, and normal operations?
3.3 Research objective 3
To empirically assess the influence of countermeasures on human behavior during active
shooter incidents in office and school buildings.
• Research Question 3.1: How does the implementation of countermeasures influence
human behavior during active shooter incidents in office and school buildings?
• Research Question 3.2: How do the building and social contexts (office vs. school) affect
human behavior during active shooter incidents?
• Research Question 3.3: How do human behaviors vary by people’s occupational
backgrounds during active shooter incidents?
3.4 Research objective 4
To leverage empirical knowledge about human-building-emergency interactions to
develop data-driven crowd evacuation simulations for informing building safety design.
• Research Question 4.1: How to quantify human-building-emergency interactions and
represent them in agent-based crowd evacuation simulations?
• Research Question 4.2: How to develop crowd evacuation simulations and employ them to
evaluate building safety designs?

29

Chapter 4. Influence of Architectural Visual Access on Wayfinding Behavior in
Building Fires
In this study, we focused on how architectural visual access (thereafter referred to as visual
access) resulting from building design strategies affect people’s wayfinding behavior in building
fires. We also explored how the tendency of following or avoiding the crowd during the evacuation
process differ under different visual access conditions. In addition, we investigated the impact of
participants’ cultural backgrounds on their wayfinding behavior in building fires. The balance of
this chapter is organized as follows. In section 4.1, the methodology of this study, including the
experiment design, recruitment of participants, VR apparatus and simulations, experiment
procedure, and data analysis is presented. The results of this study are reported in section 4.2.
Discussions around the results, limitations of this study, and directions for future research are
presented in section 4.3. Finally, section 4.4 provides a summary of this study and concludes the
chapter.
4.1 Methodology
4.1.1 Experiment design
To study the influence of visual access on emergency wayfinding, we examined
participants’ evacuation behavior through a hypothetical evacuation scenario due to a train fire in
a virtual metro station. Since people dynamically make decisions and adjust their evacuation
strategies during building emergencies [94], simplified indoor environments such as a single room
with only one decision point may not precisely reflect people’s wayfinding behavior. Therefore, a
virtual metro station was modeled, which was based on an existing metro station in Beijing, China,
as shown in Fig. 1.
30

Fig. 1. Comparisons of the real and virtual metro station (left: real metro station, right: virtual
metro station).
The metro station consists of two floors: the ground floor and the underground floor, as
shown in Fig. 2. There are two platforms on the two sides of the railway (shown in dark grey in
Fig. 2) on the ground floor. Moreover, there are six staircases and escalators connecting the two
floors. As each staircase is paired with an escalator (shown in Fig. 1), a pair of staircase and
escalator is denoted as staircase (e.g., Staircase 1) thereafter. Additionally, the metro station has
three exits, all of which are on the ground floor. On the ground floor, Hallway 1 (part of the
platform) is approximately 22 m, Hallway 2 is approximately 64 m, and the segment from the end
of Hallway 2 to Exit 1 is approximately 28 m long. On the underground floor, the ticket lobby
(surrounded by the lower end of the six staircases) has an area of approximately 25 m × 25 m.

Fig. 2. Illustration of the metro station layout, decision points and evacuation routes.
31

The virtual fire modeled in this study was based on a fire accident happened at the Tsim
Sha Tsui station in Hong Kong on February 10, 2017 [163]. While we did not intend to exactly
replicate this incident in the present study, we used it as a reference to develop the following
evacuation scenario: The virtual fire initially broke out in the second compartment of the train
(consisted of six compartments) approaching the metro station, then it spread to other
compartments. When the train stopped and the doors opened, the smoke further spread to the metro
station, as shown in Fig. 3. An emergency announcement was broadcasted in both Chinese and
English. The starting point of the evacuation was set to be at the midpoint of the platform on the
ground floor, as shown in Fig. 2. Participants were set to face the railway at the starting point and
the train on fire approached the participants from the left side. At this point, participants had to
make a directional choice, thus the starting point was also denoted as Decision Point 1 (DP 1):
they could either take Staircase 1 to go to the underground floor (i.e., Route 4 or 5) or go to Hallway
1 on the other side of DP 1 (i.e., Routes 1, 2, or 3). If participants chose Route 1, 2 or 3, they had
to make another directional choice after they arrived at the intersection of Hallways 1 and 2,
marked as DP 2 in Fig. 2. At this point, participants could choose to keep going forward via
Hallway 2 and evacuate the station via Exit 1 (invisible from DP 2) on the ground floor (i.e., Route
2), or they could go downstairs using Staircase 2 in Hallway 2 (i.e., Route 1 or 3). Once participants
chose either Route 1 or 3, they would be on the underground floor and had to navigate to DP 3 and
use one of the two staircases (i.e., Staircase 3 or 4) to evacuate the station via Exit 2 or 3. If
participants chose Route 4 or 5 at DP 1 and went to the underground floor via Staircase 1, likewise,
they would need to move to DP 3 and choose Staircase 3 or 4 to evacuate the metro station.
Therefore, as discussed above, there are five possible evacuation routes in total, which are
presented in Fig. 2. The approximate length of each route is as follow: Routes 1 and 3: 111 m,
32

Route 2: 103 m, and Routes 4 and 5: 105 m. At each DP, both routes are marked with signage,
which has arrow pointing to a direction along with corresponding words (e.g., exit) in both Chinese
and English, hence can be understood by participants in all three locations.

Fig. 3. Illustration of the virtual fire and smoke.
Additionally, it is worthwhile to point out that while at all of the three DPs, participants
needed to choose from two available directions, the decision-making conditions at each DP were
different: (1) For DP 1, participants needed to make an immediate decision after being immersed
in the IVE and perceiving the fire, without any further exploration of the environment; (2) For DP
2, participants needed to decide after entering Hallway 2 from Hallway 1. Thus, they had more
time to observe the environment compared with DP 1, but they still needed to make a quick
decision since DP 2 was close to the intersection of Hallways 1 and 2; (3) For DP 3, when
participants reached the underground floor, they needed to travel a relatively long distance to arrive
at DP 3 and make the final directional choice. Thus, compared with DPs 1 and 2, participants had
more time when moving to DP 3 to perceive the environment and adjust their decisions. Inclusion
of several DPs with varying conditions is a unique contribution of this study.
Visual access in the station was manipulated by several design strategies and resulted in
another version of the station, as shown in Fig. 4 – 6 (details are shown in the zoomed-in images).
First, since Staircase 1 near the platform was not visible at DP 1, the solid wall next to Staircase 1
was replaced by a glass wall (marked with a red rectangle in Fig. 4 (b)), so that participants could
see Staircase 1 through the glass wall. Second, at DP 2, the columns in Hallway 2 were removed
33

to improve the visual access (Fig. 5). To further increase the visual access of Hallway 2, the solid
wall along the right side of Hallway 2 was also replaced by a glass wall so that participants could
see the sky and the outdoor environment through the glass wall, and the ticket booths were moved
to make them visible from DP 2. Likewise, on the underground floor, columns were removed and
the solid wall next to Staircase 4 was replaced by a glass wall so that participants could see
Staircase 4 through the glass wall (Fig. 6). Ticket vending machines by the wall were also moved
not to block participants’ sight through the glass wall. One important note is that while two
directions were available at each DP, the visual access was mainly improved for one of the
directional choices (i.e., Staircase 1 at DP 1; Hallway 2 at DP 2; and Staircase 4 at DP 3). The
reason was to examine whether the increased visual access of a direction would encourage people
to choose it during emergency evacuation.

(a) Low visual access (b) High visual access
Fig. 4. Design strategy taken to manipulate visual access at DP 1.
34

(a) Low visual access (b) High visual access
Fig. 5. Design strategy taken to manipulate visual access at DP 2.

(a) Low visual access (b) High visual access
Fig. 6. Design strategy taken to manipulate visual access at DP 3.
Other evacuees were also included in the IVE by incorporating non-player characters
(NPCs). In total, fifty-three NPCs were included. The NPCs varied in their gender, age and
appearance, and moved at a constant speed varying from 0.7 m/s to 2.8 m/s, based on the Chinese
national code for metro safety evacuation [164]. The NPCs were positioned at pre-determined
locations at the metro platform (46 out of 53) or in the train compartments (7 out of 53) at the
beginning of the experiment. The NPCs were set to have a view angle of 120 degrees. They had
idle animation (looking around) when standing at their initial locations, in order not to have them
stand completely still and look unrealistic. NPCs at the platform started to evacuate once the fire
and smoke were within their view angle with a distance less than 15 m. NPCs in the train started
to evacuate once the train fully stopped and the doors opened. During the evacuation process, the
35

NPCs ran to a series of predetermined locations following their assigned evacuation routes, until
their final destinations (i.e., exits) were reached. To investigate how visual access influences
wayfinding behavior and to isolate the impact of crowd, in one condition, NPCs were set to split
almost evenly between two available directions at each DP. To examine how different visual
access levels would affect the tendency of following or avoiding the crowd, in the second condition,
NPCs were set to split unevenly (approximately 80% vs. 20%) between the two directions at each
DP. For the second condition of NPCs’ evacuation process, it is important to mention that, in order
to examine if evacuees would avoid the crowd when alternative direction with high visual access
was available, at each DP, the majority of the NPCs were set to choose the direction that was not
made more visible in the high visual access condition, while the minority of the NPCs chose the
direction that was made more visible in the high visual access condition. NPCs’ directional choices
at each DP and route choices are summarized in Table 1 and Table 2, shown in both numbers and
approximate percentages.
Table 1. Directional choices of NPCs at each decision point.
Directional
choices
DP 1 DP 2 DP 3
Hallway 1 Staircase 1 Staircase 2 Hallway 2 Staircase 3 Staircase 4
NPCs evenly
distributed
27 (51%) 26 (49%) 14 (52%) 13 (48%) 20 (50%) 20 (50%)
NPCs
unevenly
distributed
43 (81%) 10 (19%) 34 (79%) 9 (21%) 35 (80%) 9 (20%)
Table 2. Route choices of NPCs during evacuation process.
Route choices Route 1 Route 2 Route 3 Route 4 Route 5
NPCs evenly
distributed
7 (14%) 13 (24%) 7 (14%) 13 (24%) 13 (24%)
NPCs unevenly
distributed
27 (51%) 9 (17%) 7 (13%) 8 (15%) 2 (4%)
Thus, there were four experimental scenarios in total: (1) Scenario A: low visual access
with even distribution of NPCs; (2) Scenario B: high visual access with even distribution of NPCs;
(3) Scenario C: low visual access with uneven distribution of NPCs; and (4) Scenario D: high
36

visual access with uneven distribution of NPCs.
4.1.2 Participants
To investigate the impact of cultural background on emergency wayfinding, the data
collection was carried out in three different locations, including London, U.K., Beijing, China and
Los Angeles (LA), U.S. These three locations were selected for data collection because: first,
London, Beijing, and LA are three cities in three different continents, representing three distinct
locations; second, compared with American and British cultures, Chinese culture is considered to
have more collectivism than individualism [165], which might affect people’s following or
avoiding tendency during evacuations; third, London, Beijing and LA all have metro systems.
However, the metro systems in these three cities vary in their scale and ridership: metro systems
in Beijing have the largest size and ridership, followed by London, and then LA [166,167].
This study was approved by the University Park Institutional Review Board (UPIRB) of
University of Southern California. Emails, flyers, personal solicitation, and outlets on social media
were used to recruit participants. Participants in Beijing received 30 CNY as monetary incentives,
while those participated in London and LA did not receive any compensation. To participate in the
experiment, participants had to meet several criteria, including: (1) no heart-related illness, (2) no
wrist/hand injuries, (3) no previous uncomfortable VR experience, and (4) normal or corrected-to-
normal vision. In total, 226 participants (66 in London, 83 in Beijing, and 77 in LA) were recruited
in this study. Based on the Scenarios A – D described in section 4.1.1, participants were divided
into Groups A – D depending on their assignments to the scenarios. Sample size and demographic
information of the participants, including their locations, age and gender, in each of the
experimental groups are shown in Table 3.
Table 3. Sample size and demographics of the experimental groups.
37

Group Sample size
Age
Location Male Female
M SD
Group A 59 25.5 6.1 London 9 9

Beijing 12 10
LA 9 10
Group B 56 24.1 5.9 London 9 7

Beijing 11 9
LA 10 10
Group C 55 25.7 8.3 London 8 7

Beijing 11 10
LA 9 10
Group D 56 23.4 5.6 London 9 8

Beijing 11 9
LA 9 10
4.1.3 VR apparatus and simulations
The equipment used in the experiments included two computer workstations and an HTC
Vive VR system. The two computer workstations were connected in a local area network, using
Photon Server software. One computer workstation was used as the client and was connected to
the HTC Vive VR system to run the VR experiment; the other computer workstation worked as a
server that controlled the execution of the VR experiment and recorded the data (i.e., participants’
evacuation time, distance and trajectory) during the experiment [168]. The HTC Vive VR system
included a head-mounted-display (HMD), which was used for the visual display of the IVE, a
controller for self-navigation in the IVE at a constant moving speed of 2.4 m/s (which was decided
during pilot studies based on speed’s contribution to motion sickness in VR), two base stations for
positioning the HMD and the controller, and a headphone connected to the HMD to provide audio
stimuli (e.g., emergency broadcasting, fire alarm, etc.). Participants could change their orientation
in the IVE by changing their head orientation in the physical world and they could move in the
IVE by using the controller.
3D Studio Max software was used to model and render the virtual metro station. Then, the
model was imported to Unity3D game engine to create the fire emergency scenario using the
38

embedded particle system in Unity3D. The duration, size, and speed of the virtual fire and smoke
were manually predefined in the embedded particle system in Unity3D, to make the virtual fire
and smoke visible and look realistic in the IVEs. NPCs were modeled in 3D Studio Max software,
and their evacuation routes were preprogrammed in Unity3D. Participants’ interactions with the
environment (e.g., navigating in the station) were also incorporated in Unity3D. The location of
the participants in the metro station were updated and recorded per second during the experiment.
4.1.4 Procedure
Prior to the experiment, participants read and signed an IRB-approved consent form, which
described that the aim of the study was to investigate how building design and social interactions
would influence people’s responses during emergencies. Participants were asked to complete a
screening survey, which included questions related to their basic health conditions to determine
their eligibility for participation. Participants who did not meet all of the criteria were thanked and
dismissed from the experiment. If considered eligible, participants were allowed to proceed to
complete a pre-experiment questionnaire, which asked their basic demographic information (e.g.,
gender, age, nationality, current country of residence, etc.), their positive and negative emotions
measured with the Positive Affect and Negative Affect Scale (PANAS) [169,170], and their
simulator sickness measured with the Simulator Sickness Questionnaire (SSQ) [171].
Upon completion of the pre-experiment survey, participants were instructed to put on the
HMD and went through training to practice operations in an IVE. It is important to note that the
training environment, which was an empty open space, was different from the experimental
environment. Once participants felt familiar with the VR operations, they were asked to take off
the HMD and read the experiment instructions, which informed the participants that their task was
to find a way to evacuate the station. Once participants finished reading the instruction and
39

obtained any necessary clarification from the experimenter, they were randomly assigned to one
of the four experimental groups. During the experiment, once participants reached any of the exits,
they were asked to take off the HMD and continue to complete a post-experiment survey. In the
post-experiment survey, information collected from participants’ responses included: (1) the
importance ranking of different factors (i.e., visibility of exits, ticket booths, staircases, distance
to fire, and directions indicated by the crowd flow and signage) for their directional choices at each
DP; (2) the ratings of the importance of visual access and crowd flow on their directional choices
at each DP; (3) their positive and negative emotions measured with PANAS [169,170]; (4) their
simulator sickness measured with SSQ [171]; (5) their sense of direction measured with the Santa
Barbara Sense of Direction Scale (SBSOD) [172]; (6) their sense of presence in the IVE measured
with the presence questionnaire (PQ) [173]; (7) their level of wayfinding anxiety measured with
the Lawton’s spatial anxiety scale [174]; and (8) their past experiences of evacuation in building
emergencies. Each participant only took part in the experiment once. After completing the post-
experiment survey, participants were thanked and dismissed.
4.1.5 Data analysis
Chi-square test was used to analyze how visual access and culture affected participants’
wayfinding behavior, as well as if there was any group difference because of its capability of
describing the relationship between two nominal variables. When Chi-square could not be used
due to insufficient sample size, Fisher’s exact test was used instead, to compare participants’ route
and directional choices in different groups. Independent samples t-test was used for between-group
comparisons, including participants’ evacuation performance, their evaluation of various factors
that influenced wayfinding, as well as whether there was any difference in age, sense of presence,
sense of direction etc. Additionally, one sample t-test was used to analyze the change of
40

participants’ emotions and simulator sickness during the experiment. Shapiro-Wilk test was used
to examine whether the data was normally distributed. If the normality requirements were not
satisfied for parametric statistical tests (independent samples t-test and one sample t-test),
nonparametric statistical tests, Mann-Whitney U test and Wilcoxon Signed Rank Test were used
for between-group comparisons and within-group comparisons, respectively. The significance
level was set as 0.05 and marginal significance level was set as 0.10. All data analysis was
conducted using SPSS 25.
4.2 Results
Prior to analyzing the results, the comparison between Groups A and B, Groups C and D
was conducted, in terms of participants’ age, gender, education level, sense of direction,
wayfinding anxiety, sense of presence, change in simulator sickness and change in emotions during
the experiment. The results showed that there was no difference between Groups A and B, and
between Groups C and D, in any of the above-mentioned measures, which eliminated the possible
influence of these factors on the difference between the groups.
It was found that after the experiment, participants’ ratings of being enthusiastic (z = -5.059,
p < 0.001), determined (z = -2.107, p = 0.035), and nervous (z = -3.030, p = 0.002) significantly
decreased, and being alert (z = 3.673, p < 0.001), distressed (z = 2.093, p = 0.036), and scared (z
= 2.001, p = 0.045) significantly increased. The results showed the virtual fire emergency did in
fact evoke participants’ emotional arousals. Moreover, participants’ responses to the presence
questionnaire reported an average PQ score of 141.81 (SD = 17.57) from the range of 30 (no
presence) to 210 (presence as reality). Compared with the PQ score in prior studies (e.g., mean =
98.11, SD = 15.78 in [173], and mean = 90.30, SD = 14.5 in [175]), our results suggested that the
inclusion of NPCs, visual (e.g., virtual fire and smoke) and audio stimuli (e.g., emergency
41

broadcasting and alarm) did impose an adequate sense of emergency on participants.
4.2.1 Influence of visual access with even distribution of NPCs (Groups A and B)
Whether the low and high visual access influenced the participants’ route choices in Groups
A and B were analyzed first. Fisher’s exact test was used for analyzing the participants’ route
choices, which revealed that overall, participants’ choices of the 5 evacuation routes were not
significantly different between Groups A and B (p = 0.428 > 0.10), indicating that visual access
did not affect participants’ choices among the 5 evacuation routes, as shown in Fig. 7. While the
overall patterns of route choices were similar between the two groups, it is critical to further
investigate the participants’ directional choices at each DP, as the participants’ decisions at DPs
fundamentally determined their evacuation trajectories and the visual access was manipulated at
the DP level, not at the route level.

Fig. 7. Route choices of participants in Groups A and B and NPCs.
Directional choices of participants in London, Beijing and LA at DP 1 were compared
between Groups A and B. Chi-square test showed that in Group A, the directional choices of
London, Beijing, and LA participants at DP 1 did not have any significant difference (χ² (2, 59) =
42

0.245, p = 0.885). However, the results revealed that in Group B, the directional choices of
participants from the three locations were significantly different at DP 1 (χ² (2, 56) = 9.333, p =
0.009), as shown in Fig. 8. Compared with Beijing and LA participants, more London participants
chose to go to Staircase 1 (with improved visual access).

(a) Group A (b) Group B
Fig. 8. Directional choices of London, Beijing, and LA participants at DP 1 (* denotes the
direction with improved visual access).
To further explore how various factors influenced directional choices at DP 1, participants’
subjective evaluations of these factors were analyzed. The results showed that at DP 1, there was
no significant difference in the evaluation of these factors between the two groups (all p > 0.10).
Additionally, to investigate the difference between London participants and Beijing/LA
participants at DP 1, the subjective evaluation of London participants who took Staircase 1 in
Group B was compared with those of Beijing and LA participants who did not take Staircase 1 in
Group B. The results of Mann-Whitney U test revealed that London participants who took
Staircase 1 in Group B considered that ‘visibility of staircase’ was significantly more important in
their directional choice at DP 1 than those of Beijing and LA participants who did not take
Staircase 1 in Group B (U = 253, z = 2.647, p = 0.009). To further look into the influence of
‘visibility of staircase’, the subjective evaluation of London participants who took Staircase 1 in
Group B was compared with those of Beijing and LA participants who took Staircase 1 in Group
43

B as well. The results showed that the ranking of ‘visibility of staircase’ was not significantly
different (U = 67, z = 0.878, p = 0.426). The above results indicated that at DP 1, participants who
took Staircase 1 in Group B was indeed because of its improved visual access.
Chi-square test showed that overall, visual access had a marginally significant effect on
participants’ directional choices at DP 2 (χ² (1, 72) = 2.794, p = 0.095). Compared with the
participants in Group A, more participants in Group B went to Hallway 2, which was made more
visible in Group B by removing the columns, relocating ticket booths and changing the wall
material in Hallway 2. Moreover, participants’ higher preference of Hallway 2 in Group B was
consistent in London, Beijing and LA, as shown in Fig. 9.

(a) Group A (b) Group B
Fig. 9. Directional choices of London, Beijing, and LA participants at DP 2 (* denotes the
direction with improved visual access).
Participants’ evaluation of influencing factors on their directional choices at DP 2 was also
analyzed. The results showed that there was a significant difference in the participants’ evaluation
of ‘direction indicated by signage’ between Groups A and B (U = 837, z = 2.173, p = 0.030).
Participants in Group A considered the direction indicated by signage more important in their
decision making, compared with those in Group B.
With respect to DP 3, participants could arrive at this location via two possible routes: (1)
go to Hallway 1 at DP1, then take Staircase 2 at DP 2 to go to the underground floor and reach DP
44

3; or (2) go to Staircase 1 at DP 1 and then navigate to DP 3 on the underground floor. Regardless
of the way they reached DP 3, all participants were combined together to analyze their directional
choices at DP 3. The result of the Chi-square analysis showed that participants’ directional choices
were not significantly different at DP 3 (all p > 0.1) between Groups A and B or among the three
locations. Moreover, the subjective evaluation of the influencing factors was not significantly
different between the two groups.
Three measures, namely virtual evacuation time, distance and speed, were used to assess
participants’ virtual evacuation performance. Evacuation time is a critical parameter in
emergencies [176], evacuation distance is related to people’s travel paths, which are also an
important factor for people’s safety [177], and speed describes the relationship between time and
distance. In this study, virtual evacuation time was defined as the time that a participant spent from
hearing the fire alarm to arriving at one of the exits, which is consistent with the definition of
evacuation time in the literature (i.e., sum of the time to receive warning, time to respond to
warning, delay time, and movement time) [178,179]. Virtual speed was defined as a participant’s
total movement distance divided by his/her total evacuation time, which was affected by
participants’ stoppage during the evacuation to make directional choices, the time participants
spent before starting evacuation, etc.
To examine how participants’ virtual evacuation performance differed in the two groups
and to eliminate the influence of different length of each route, an analysis of participants’ virtual
evacuation performance was conducted for each individual route. The only difference between
Routes 1 and 3 or Routes 4 and 5 was the directional choice at DP 3, which did not impact the
overall route distance. If analyzed separately, the sample size for each route would be too small to
conduct the statistical test, hence Routes 1 and 3 as well as Routes 4 and 5 were combined for the
45

analysis. The results are presented in Table 4. It was revealed that visual access had a significant
effect on the virtual evacuation time of participants who took Routes 1 and 3, Route 2, and Routes
4 and 5 (all p < 0.05). Additionally, participants who took Routes 1 and 3 (p = 0.005) and Route 2
(p < 0.001) in Group B had significantly higher virtual evacuation speed compared with those in
Group A. Moreover, it was found that participants who took Routes 4 and 5 in Group B travelled
significantly less distance compared those in Group A (p = 0.021).
Table 4. Comparison of virtual evacuation performance for each route between Groups A and B.
** p < 0.05. * p < 0.1 (+ denotes Mann-Whitney U test was used for the comparison).
Virtual
evacuation
performance
Routes
Low visual access High visual access
P
M SD M SD
Time (s) Routes 1 and 3 83.5 13.0 68.1 6.1 0.047 **
Route 2 65.8 7.7 60.4 7.7 0.001 ** (+)
Routes 4 and 5 75.9 12.7 68.8 10.8 0.017 ** (+)
Distance (m) Routes 1 and 3 133.5 13.3 129.4 5.1 0.839 (+)
Route 2 113.7 7.4 112.3 6.9 0.400 (+)
Routes 4 and 5 121.1 7.9 116.4 11.5 0.021 ** (+)
Speed (m/s) Routes 1 and 3 1.6 0.2 1.9 0.1 0.005 **
Route 2 1.7 0.2 1.9 0.2 < 0.001 ** (+)
Routes 4 and 5 1.6 0.2 1.7 0.2 0.257
4.2.2 Influence of visual access with uneven distribution of NPCs (Groups C and D)
Participants’ route choices in Groups C and D were first analyzed to evaluate whether
manipulation of visual access influenced their following or avoiding tendency during the
evacuation process. Fisher’s exact test showed that the choice of the 5 possible evacuation routes
varied significantly between Groups C and D (p = 0.015). As shown in Fig. 10, in Group C, the
most frequently chosen route was Route 1, which was also taken by the large majority of the NPCs.
Nevertheless, in Group D, Route 2, which was made more visible in the experiment, was the most
frequently chosen even though it was taken by only 9 of the 53 NPCs.
46

Fig. 10. Route choices of participants in Groups C and D and NPCs.
To further analyze the influence of visual access on participants’ following and avoiding
tendency, their directional choices at each DP were compared. Participants’ directional choices at
DP 1 in both Groups C and D are shown in Fig. 11, and there was no significant difference among
London, Beijing and LA participants (all p > 0.10). This result indicated that participants followed
the crowd in both conditions. This result was also correlated by participants’ evaluation of the
influencing factors, which showed that crowd flow was the most influential factor and there was
no significant difference in the evaluation of other factors between the two groups (all p > 0.10).

Fig. 11. Directional choices of participants at DP 1 in Groups C and D (* denotes the direction
47

with improved visual access in Group D, + denotes the direction that the majority of NPCs took).
Directional choices of participants at DP 2 in Groups C and D are shown in Fig. 12. It was
found that unlike London and LA participants who tended to choose Hallway 2 (which was made
more visible) in Group D, Beijing participants’ directional choices did not vary between the two
groups: the majority of Beijing participants (around 63%) took Staircase 2 in both groups.
Additionally, in their subjective evaluation, Beijing participants in both groups agreed that they
considered the crowd as an important factor and the evaluation of the importance of crowd was
consistent in both groups (p > 0.10).

(a) Group C (b) Group D
Fig. 12. Directional choices of participants at DP 2 (+ denotes the direction that the majority of
NPCs took; * denotes the direction with improved visual access in Group D).
Additionally, Fig. 13 shows participants’ directional choices at DP 3, and there was no
significant difference among participants in London, Beijing and LA (all p > 0.10). Fisher’s exact
test showed that there was marginally significant difference in the directional choices between
Groups C and D (p = 0.057). The analysis indicated that most participants followed the crowd at
DP 3 when the visual access was low. On the contrary, when the alternative direction was made
more visible, more participants, compared with those in Group C, chose to evacuate via the more
visible direction. However, the analysis of participants’ subjective evaluation of the influencing
factors did not reveal any significant difference at DP 3 between the two groups (all p > 0.10).
48

Fig. 13. Directional choices of participants at DP 3 in Groups C and D (* denotes the direction
with improved visual access in Group D, + denotes the direction that the majority of NPCs took).
To further examine how visual access influenced participants’ virtual evacuation
performance with uneven distribution of NPCs, participants’ virtual evacuation time, distance and
speed were compared and analyzed, as shown in Table 5. Since very few participants in Groups C
and D took Routes 4 and 5 (3 in Group C and 3 in Group D), no statistically reliable conclusions
could be drawn, hence these two routes were not included in Table 5 and the analysis.
Table 5. Comparison of virtual evacuation performance for each route between Groups C and D.
** p < 0.05. * p < 0.1 (+ denotes Mann-Whitney U test was used for the comparison).
Virtual
evacuation
performance
Routes
Low visual access High visual access
p
M SD M SD
Time (s) Routes 1 and 3 80.6 11.2 80.1 13.4 0.629 (+)
Route 2 74.1 16.8 62.8 9.5 0.005 ** (+)
Distance (m) Routes 1 and 3 134.1 11.8 135.2 9.5 0.507 (+)
Route 2 116.4 14.0 115.2 13.5 0.831 (+)
Speed (m/s) Routes 1 and 3 1.7 0.2 1.7 0.2 0.505
Route 2 1.6 0.3 1.9 0.2 0.002 **
It was found that visual access did not have a significant effect on the virtual evacuation
performance of participants who took Routes 1 and 3 (all p > 0.10). On the contrary, for Route 2,
which was the route taken by minority of the crowd, there existed significant difference in
participants’ virtual evacuation time (p = 0.005) and speed (p = 0.002).
4.2.3 Interactive influence of visual access and crowd flow (Groups A, B, C and D)
The above results suggest that both visual access and crowd flow could influence
49

participants’ wayfinding behavior, hence their interactive influence is considered as well. Fig. 14
shows the participants’ evacuation trajectories. With the same level of visual access, crowd flow
affected participants’ evacuation by causing more participants to follow the crowd under the
uneven distribution of NPCs (comparing the green rectangle areas between Fig. 14 (a) and (c), (b)
and (d), (e) and (g), and (f) and (h)). However, comparing the red rectangle areas between Fig. 14
(c) and (d), (g) and (h), it was also illustrated that when the level of visual access was high,
relatively fewer participants took the evacuation route that was taken by the majority of NPCs.
Additionally, the comparison of evacuation time revealed that participants who took Route 2 in
Group C spent marginally longer time than those who took Route 2 in Group A (U = 292.5, z =
1.923, p = 0.054), whereas there was no such difference in the high visual access condition (p >
0.10).

(a) Group A, ground floor (b) Group B, ground floor

(c) Group C, ground floor (d) Group D, ground floor
50

(e) Group A, underground floor (f) Group B, underground floor

(g) Group C, underground floor (h) Group D, underground floor
Fig. 14. Participants’ evacuation trajectories.
4.3 Discussion
4.3.1 Influence of visual access on directional choices during building emergencies
Except for the London participants in Group B, visual access did not influence participants’
directional choices at DP 1. One reason might have been related to the participants’ stress level
when encountering an emergency and unfamiliarity with the environment, which could narrow
their attention [180] and impair their ability to process environmental information [181]. At DP 1,
participants needed to make a directional choice immediately after they were immersed in an
unfamiliar emergency environment, as a result, improved visual access may not be perceived when
participants were highly stressed. Another reason might have been the location of building
elements as well as the hazards (i.e., fire in this study). At DP 1, Staircase 1 was located in the
direction where the train on fire approached from. It was suggested in the literature that the more
people are unfamiliar with the environment, the more likely they would choose to stay away from
the dangerous area instead of going towards it [182]. Thus, in the high visual access condition,
51

more participants chose to go to Hallway 1 instead of Staircase 1, to avoid getting close to the fire.
Participants’ evaluation of the influencing factors at DP 2 showed that participants in
Group A considered the direction indicated by signage more important in their decision making,
compared with those in Group B. This result might also be related to the building characteristics
and design strategies used to manipulate visual access at DP 2. When visual access was low,
participants’ line of sight was partially blocked by columns and solid walls in Hallway 2. Therefore,
the signage in Hallway 2 and near Staircase 2, which was at the ceiling level and hence less
influenced by the manipulation of visual access, might have been more influential in participants’
decision making. This finding is in line with literature: people are more likely to choose the visible
direction instead of heading towards another direction that is unknown to them [183].
At DP 3, visual access did not significantly influence participants’ directional choices. One
factor that distinguishes DP 3 from DP 1 and DP 2 was that the participants had more time to make
directional choices at DP 3, compared with DPs 1 and 2. In fact, Staircases 3 and 4 were located
symmetrically in relation to DP 3. Thus, in the high visual access condition, even though Staircase
4 was more visible when participants were moving towards DP 3 on the underground floor, they
might have adjusted their final decisions when arriving at DP 3, and due to the symmetrical design
of Staircases 3 and 4, these two exits became similarly visible at DP 3.
We conclude that visual access could influence participants’ directional choices, however
this is contingent upon other contextual factors. First, the magnitude of visual access has direct
correlation with people’s choices. The clearer a direction leads to an exit, the more likely it is to
be chosen. Second, when people’s stress level is high during building emergencies, their ability to
perceive the environmental information is likely to be reduced and the effect of visual access might
be reduced accordingly. Third, spatial characteristics of the building also determine the effect of
52

visual access. If a route is intrinsically unattractive (e.g., because of its location), improving its
visual access may still not motivate people to choose it.
4.3.2 Influence of visual access on participants’ tendency of following and avoiding the crowd
At DP 1, the participants in Groups C and D followed the crowd regardless of the visual
access level. Majority of the NPCs that moved to Hallway 1 provided strong directional
information and caused participants to go to Hallway 1 instead of Staircase 1. When people are
stressed and unfamiliar with the environment, as the participants likely experienced at DP 1, they
would tend to follow the crowd during evacuation [184]. At DP 2, visual access persuaded more
participants to choose alternative visible direction rather than solely following the crowd. This
finding is in agreement with prior studies [183]. It was also reported in prior studies that if the
crowd is moving to an exit that is invisible to evacuees, they may think others know something
that they do not, and their tendency of avoiding the crowd is reduced [55]. In Group D, the visual
access of Hallway 2 was significantly improved by removing columns, changing wall materials
and relocating ticket booths, which provided very strong directional information and might
outweighed the effect of crowd flow. Additionally, at DP 3, more participants in Group D chose
the more visible direction compared with those in Group C. As discussed above, as participants
could see both Staircases 3 and 4 lead to outside, some participants decided not to follow the crowd
and took the alternative route instead.
In summary, visual access could influence people’s following/avoiding behavior and
motivate people to move towards more visible directions. This could have important practical
implications, such as designing buildings for more efficient evacuation and better estimating
required safe egress time in performance-based design.
4.3.3 Influence of visual access on virtual evacuation performance
53

By comparing participants’ evacuation performance between Groups A and B, visual
access was found to be positively related to participants’ virtual evacuation performance.
Evacuation process consists of three phases: (1) awareness of danger by external stimuli (cue
validation), (2) validation of and response to environmental factors (decision-making), and (3)
movement to/refuge in a safe place (movement/refuge) [102]. Apart from going through these
phases at the beginning of the evacuation process, people dynamically determine their final
destinations during the evacuation process and consistently experience the following phases:
perceiving the environment, making decisions, and evacuating [94]. Therefore, improvement of
visual access could enhance the participants’ evacuation performance in two ways. First, improved
visual access could reduce the level of uncertainty perceived by participants by providing more
visual information about the environment. Since uncertainty is an important factor that prolongs
the decision-making phase, the reduction of uncertainty could enable the participants to make
quicker decisions during the evacuation process [107]. Second, the removal of obstacles could
further facilitate participants’ evacuation in the movement phase.
Comparing participants’ virtual evacuation performances between Groups C and D, it was
found that visual access did not have a significant effect on the virtual evacuation performance of
participants who took Routes 1 and 3. One major reason that may have contributed to this result
had to do with the crowd, since the large majority of the crowd followed Routes 1 and 3 until their
final route choices at DP 3. This suggested that, while visual access was different between Groups
C and D, the crowd was the most decisive factor for the participants who took Routes 1 and 3 in
Groups C and D. However, it was also found that for participants who took Route 2 in Groups C
and D, even though their virtual evacuation distance was similar between the two groups, high
visual access still helped participants to improve their virtual evacuation performance.
54

4.3.4 The interaction effect of visual access and crowd flow on participants’ wayfinding
behavior
Two reasons might have caused the participants to follow the crowd with the same level of
visual access. First, during building emergencies, especially in an unfamiliar environment,
following behavior is common to occur [13]. Second, crowd flow conveys directional information
that might be easier to be perceived compared to the static information (e.g., signage, visual access)
[185], hence imposing major influence on participants’ wayfinding behavior during evacuation.
Additionally, participants in Group C who took Route 2 had longer virtual evacuation time
compared with those in Group A. This result was probably due to the fact that in the low visual
access condition, the participants who took Route 2 (avoiding the crowd at DP 2) spent more time
making this decision as there was more uncertainty in the environment. Therefore, it prolonged
the process to make the decision of avoiding the crowd. In the high visual access condition,
however, the improvement of visual access made Hallway 2 more visible, which shortened the
decision-making process for the participants to choose Hallway 2. Hence, visual access and crowd
flow are indeed two interrelated factors and their influence on wayfinding behavior depends on
one another. When applying the findings on visual access and crowd flow, they should be
considered collectively instead of independently.
4.3.5 The cultural impact on participants’ wayfinding behavior
Overall, London, Beijing, and LA participants had similar evacuation route choices, with
a few exceptions at certain DPs. First, with even distribution of NPCs, the visual access
improvement of Staircase 1 was perceived by more London participants compared with Beijing
and LA participants. This result might be related to London participants’ relatively richer prior
experience with metro stations, which could lower the stress level they experienced during the
55

experiment [186]. In fact, the experiments were conducted in university campuses in Beijing and
LA, while the data collection location in London was very close to a major connection point in the
metro system of London. Second, at DP 2, while London and LA participants tended to avoid the
crowd and went to a more visible direction in Group D, the following behavior of Beijing
participants remained at a high level even in Group D. One possible reason for this finding may
be attributed to the fact that China has a culture with lower level of individualism compared with
the U.S. and the U.K. [165], which resulted in a higher tendency of following in emergencies.
However, as presented above, London, Beijing and LA participants did not have consistent
differences during their evacuation process, therefore, whether their cultural background impacted
their evacuation behavior should be further investigated.
4.3.6 Limitations and future work
There are limitations associated with the study that require future investigations. First,
while we included virtual fire, smoke and emergency announcement in the IVEs to represent the
fire scenario, unlike real fires, the virtual fire and smoke did not affect participants’ mobility and
no thermal and olfactory stimuli were provided. Meanwhile, participants’ movement in the IVEs
was achieved by using a controller and was set at a constant speed. To enhance the sense of
presence that participants experience in the IVEs, future studies could provide more stimuli
channels (e.g., thermal, olfactory, and haptic feedback) to make the virtual fire scenario more
comparable to real fire emergencies. Second, while this study is one of the few that collected data
in multiple countries, the findings in this study were based on data collected from participants in
three locations only, thus validity of the results on people from other cultural backgrounds as well
as other factors such as, elders, children would require further investigation. Third, the results of
this study suggest that both participants’ stress level and familiarity with the building could affect
56

people’s emergency wayfinding. However, to further evaluate these effects, future studies could
monitor participants’ stress levels (e.g., using physiological measurements [187]) during the
experiment and integrate different levels of familiarity with the built environment into the design
of experiments. Finally, this study was conducted in a virtual metro station with 53 NPCs, where
no congestion was caused by the crowd. As spatial characteristics and level of crowdedness both
play a significant role in the influence of visual access, whether people would behave consistently
during emergencies in other environments, including different types of indoor spaces, such as
educational buildings, office buildings and museums, could be studied in future research.
4.4 Conclusion
In this study, we conducted a fire evacuation experiment in a virtual metro station in three
different locations (i.e., London, Beijing and Los Angeles) to understand how visual access and
people’s cultural background influence their wayfinding behavior during building fires by
comparing two levels of visual access in two different crowd conditions. There were three points
in the metro station where the participants needed to make directional choices, and multiple design
strategies (e.g., changing wall materials, removal of columns, relocating ticket booths) were used
to manipulate visual access, which, to the best of our knowledge, have not been explored in prior
studies. Participants’ route and directional choices, evacuation performance (i.e., evacuation
distance, time, and speed) and subjective measurements (e.g., emotional responses, simulator
sickness, sense of presence, wayfinding anxiety, etc.) were collected during the experiment. Our
results showed that improving visual access did attract participants to go to a more visible direction
during the evacuation, while the magnitude of the effect depended on the significance of visual
access improvement. The clearer a direction leads to an exit; the more likely participants choose
to take that direction. Moreover, improving the level of visual access could facilitate participants’
57

environmental perception and decision-making process, and encourage them to choose visible
routes over the ones taken by the crowd, although such effect may vary in different cultures and
emergency situations. The results also revealed that increasing the level of visual access in indoor
environments could improve participants’ evacuation performance during emergencies (i.e.,
shorten the evacuation distance and time, and increase the speed). Research questions 1.1 – 1.4
were addressed in this chapter.

58

Chapter 5. Preparedness of Built Environments in Response to Active Shooter
Incidents: Results of Focus Group Interviews
The present study aims to assess the effectiveness of different security countermeasures for
active shooter incidents, and how they would impact human behavior and preparedness of built
environments in response to active shooter incidents. Owing to the advantage of providing an in-
depth exploration of a topic in which many factors are still unknown [188], focus group interviews
were conducted with 15 domain experts. The balance of this chapter is organized as follows:
Section 5.1 describes the methodology of this study. The findings from the focus group interviews
are presented in section 5.2. We further discuss implications of the findings in section 5.3 and
conclude this chapter in section 5.4.
5.1 Methodology
A focus group interview refers to a discussion, in which an interviewer asks a set of open-
ended questions to a small group of target population [189]. The name is derived from the fact that
the selected groups are “focused” on a given topic [190]. Therefore, the selection criteria for
participants in focus group interviews include: having domain knowledge on the topic, being
within the age-range, and comfortable talking to the interviewer and other participants [191].
Compared to individual interviews, the main advantage of focus group interviews lies in that the
group could foster a synergy, which results in more than the sum of each individual’s output
[190,192,193]. In order to achieve the objective of this study, a series of online focus group
interviews were conducted, which could accommodate geographically distributed populations in
a synchronous virtual shared space [194]. This study was approved by the University Park
Institutional Review Board (UPIRB) of University of Southern California.
5.1.1 Participants
59

The participants of the focus group interviews were recruited through professional
organizations and societies that focus on building safety and emergency management, as well as
experts from leading engineering and design firms. Purposive sampling was used to select
participants with expertise in the two key areas, namely building design/engineering and building
security [195,196]. To be included in the study, participants must (1) be between 18 to 85 years
old, (2) be fluent in English speaking, (3) be able to hear and talk remotely, and (4) have expertise
in relevant areas, such as terrorism/active shooter incidents, school/office building design, security
engineering in buildings, and risk and emergency management. A minimum of 4 participants in
each group is generally accepted [197]. Krueger and Casey suggested between six and eight
participants in a group, but at the same time stated that a smaller group might be preferable, which
would allow participants to contribute more freely [198]. A total of 15 participants were recruited
in this study, and three focus group interviews were conducted, with each group consisting of 5
participants. The participants’ professions were security engineer (2), architect (1), emergency and
security manager (2), security consultant (5), and police officer (5). All of the participants had
professional experience and expertise closely related to active shooter incidents. Specifically, 8
participants have been involved in the law enforcement, management and training for active
shooter incidents, and the other 7 participants mainly focus on building security, including
electronics security, anti-terrorism design, and crime prevention through environmental design.
The total number of the participants (i.e., 15) and focus groups (i.e., 3) in this study was determined
by analyzing the interview results until no more themes emerged from the focus group interviews
[198].
5.1.2 Interview guide
A semi-structured interview guide was developed to direct the focus group interviews.
60

Following the suggestion in [190], the guide consists of four parts. The first part covers the
beginning of the interview, including welcome messages, explanation of the interview purpose
and use of data, obtainment of participants’ verbal consent to participate in the study and be
recorded, and self-introduction of the participants. In the second part, the interviewer asks the
participants a warm-up question: “What are the countermeasures in your experience that have been
used to proof buildings against active shooter incidents?” The purpose of this question is to
motivate the participants to share their perspectives based on their work experience and get them
involved in the discussion. Subsequently, the third part contains a series of questions to further
provoke participants’ thoughts on the topic. The purposes of these questions were to create
discussions around the effectiveness and the pros and cons of current countermeasures for active
shooter incidents. The general list of questions is illustrated in Table 6.
Table 6. General list of questions in the focus group interviews.
Questions
1. What are the countermeasures in your experience that have been used to proof buildings against active
shooter incidents?
2. How does the effectiveness of countermeasures change when the shooter is an internal threat versus
an external threat?
3. Do you see any difference in terms of these countermeasures’ implementation in different types of
buildings, for example, in school buildings vs. office buildings? If yes, how do different building types
differ in terms of countermeasures?
4. Do you have a sense of which countermeasure might affect building occupant behavior during an
active shooter incident? If so, which measures might change occupant behavior the most?
5. Do you have a sense of the effectiveness of which countermeasures might be affected by occupant
behavior during an active shooter incident? If so, the effectiveness of which countermeasure might be
changed by occupant behavior the most?
6. Are there differences or conflicts between countermeasures that try to prevent risk from other
emergencies (e.g., fires, earthquakes) in contrast to the countermeasures that are used for active shooter
incidents?
7. What are the impacts of the countermeasures for active shooter incidents on day-to-day activities?
It is important to note that while the questions listed in Table 6 acted as guidance for the
interviews, they might not be asked exactly as they are presented in Table 6. Depending on the
progress of the interview, some questions might be asked in another way or some questions might
be raised and discussed among the participants before being asked by the interviewer. Finally, the
61

last part of the interview includes summarizing the discussion and thanking the participants for
their contribution to the study.
5.1.3 Procedure
Emails were used to send invitations to potential participants, in which the purpose of the
study was mentioned. After the participants agreed to contribute to this study, the authors
forwarded a Doodle (an online calendar tool for time management and coordination of meetings)
link for the participants to provide their available timeslots. Based on the participants’ availability,
they were assigned to different groups and received BlueJeans (a cloud-based video conferencing
service) invitations. The interviews were video and audio-recorded using the built-in function of
BlueJeans, and each interview lasted between one and one-and-a-half hours. Before each interview,
the authors also sent the UPIRB-approved consent form to the participants via email, which
presented the purpose of this study, procedures, and confidentiality (i.e., the use of audio and video
dada). The interviews were conducted by two of the authors: one acted as a moderator and the
other acted as an observer [198]. The moderator led the discussion, while the observer was
responsible for taking notes and did not actively participate in the discussion. Prior to the start of
discussions, the moderator introduced the interview purpose and obtained participants’ verbal
consent to participate and be recorded during the interview, as approved by the UPIRB. The
moderator then initiated the discussion around the topic and asked questions based on the interview
guide. Upon completion of the discussion, the moderator concluded the interview and asked if the
participants had any additional thoughts or considerations. Finally, the participants were thanked
and dismissed from the interview.
5.1.4 Data analysis
Upon completion of the focus group sessions, the interviews were reviewed, and the
62

conversations were transcribed verbatim. The transcription process was based on the audio data,
while the video data was used merely to verify who was speaking in the focus group discussion.
The authors reviewed the transcripts repeatedly to get familiar with the data and obtain a sense of
interviews as a whole, which served as the preparation phase for analyzing the data [199]. Based
on the transcript, the authors used qualitative analysis to interpret the interview data [200]. The
purpose of the data analysis was to identify agreements and disputes over the topic, instead of
simply presenting numbers and percentages of participants’ responses [190,201]. Three major
stages were involved in the analysis. First, the text in the transcripts were divided into meaning
units, which could be individual word or some words in a sentence or several sentences that share
similar content (i.e., the open coding stage [202]). The meaning units were labeled with different
codes that could represent the content of the text. For instance, “One of the most impactful things
that we have been able to do and justify is reducing the number of entrances into offices.” Was
labeled as “access control.” Second, these generated codes were compared with each other and
those that were related in their content were grouped into subthemes, which is denoted as axial
coding in the literature [202]. For instance, “access control” generated in the last stage was
categorized as “countermeasures for building design and facility management.” Finally, more
abstract high-level themes were created to organize the subthemes in a hierarchical structure (i.e.,
the selective coding stage [202]). For instance, the subtheme “countermeasures for building design
and facility management” created in the last stage was organized under the theme
“countermeasures to protect buildings and occupants from active shooter incidents.” One of the
authors conducted the transcribing, coding and data analysis. The results, together with the
transcript and audio/video data were reviewed by all of the authors to verify that the outcome was
accurate and reliable. However, as only one author actually coded the transcripts, no quantitative
63

measures of agreement (e.g., interrater reliability) were included. The data analysis pipeline is
shown in Fig. 15. The findings of the interviews are presented in detail in the next section.

Fig. 15. The data analysis pipeline.
5.2 Findings of focus group interviews
The interviews were focused on evaluating the effectiveness of countermeasures to
safeguard buildings from active shooter incidents, accompanied by various considerations when
implementing these defense strategies. Through the analysis of the interviews, four themes
emerged, which are summarized along with their subthemes in Fig. 16. The themes are presented
in the following subsections, and participants’ quotes during the interviews are used for illustrative
purposes. It is important to point out that the findings are based on the participants’ perspectives
and should not be considered as factually defined.

64

Fig. 16. Themes and subthemes of the focus group interviews.
5.2.1 Countermeasures to protect buildings and occupants from active shooter incidents
With regard to protecting buildings and occupants from active shooter incidents, one of the
most commonly mentioned countermeasures by the participants was access control, such as
reducing the number of entrances, posting someone physically at the reception/security desk,
issuing proximity access control badges and possibly walk-through metal detectors. An illustrative
claim was: “If I am a bad guy, that [access control] is going to keep me from coming into the
building. [If there is strong access control], I might set up [the attack] outside the building, … but
I think that access control, in whatever it looks like for that facility, is the first step.” Some
participants pointed out that many active shooters target a soft target (i.e., a person or thing that is
relatively unprotected or vulnerable) rather than a hard target (i.e., a highly defended target), thus
the presence of access control could be a deterrent by reducing the shooter’s expectations: “They
[shooters] are not going to waste their time on trying to gain entry into an area, where they will
have a much more difficult time reaching that goal.” Nevertheless, maintaining egress during
active shooter incidents was also identified as an important factor, as an example statement being:
“There is definitely a balance between public access and security.”
Having multiple layers of security was another identified countermeasure. The primary
justification for this security design, as stated by one of the participants, is that: “You cannot ensure
that every single inch is not going to have any person that could cause harm with any weapons.”
Hence, how to draw the lines for different layers of security and build different zones is essential.
Nevertheless, concerns about implementing this countermeasure were voiced as well. A participant
stated that in a campus environment, if a shooter breaches the first line (e.g., the entrance to the
campus) and starts to attack where students are, whether or not to lock the doors of buildings where
65

more students are, would be a difficult decision. This further relates to another countermeasure
that was discussed: remote locking control mechanisms, which could secure a building or a section
of a building by activating a switch. A participant mentioned that “It [remote locking control
mechanism] is generally for an imminent threat… once the threat [a shooter] is inside, the last
thing you want to do is to lock the building, because it locks everybody [all the occupants] in.”
However, a participant commented that “When people are under stress, they might not necessarily
make rational decisions and therefore might not be able to make the right call, which might make
things worse,” referring to an accidental locking. Therefore, training for the administrators who
make these decisions was also identified as the crucial component for employing the
countermeasures correctly.
Another frequently mentioned countermeasure addressing the issue of communication was
mass notification, which could be conveyed through social media, text message, and email
notifications, etc. An illustrative comment was: “In a campus environment or an area where you
may not hear the shots being fired in one location, that [notification] is going to be very important
if someone is on the run and has not heard the gun shots.” Apart from that, some other
countermeasures were mentioned during the interviews, including using ballistic-resistant
materials, video surveillance, safe rooms, etc.
While the topic was mainly focused on countermeasures related to building design and
facility management, the participants highlighted the importance of training and regular drills. It
was mentioned that many shooters tend to surveil the location and study their target in advance,
thus the countermeasures in place cannot eliminate attacks completely. Under this circumstance,
the participants suggested that training for building occupants is a decisive factor. One example
statement was: “The fact that you have thought it [how to respond] through once is the first step,
66

and the fact you have actually done it [practice appropriate responses] once is the second step to
knowing how you should respond in an actual active shooter incident.” Similar principles were
also mentioned for law enforcement and first responder teams. Illustrative statements include: “We
let them [police agencies] conduct training in our office buildings. Therefore, when something [an
active shooter incident] does happen, it is not new to them and they have proper access to the
buildings and all [response procedures] is pre-established.” and “I always think that we could be
a bit blasé about the fact that somebody is going to make the call, and then we do not necessarily
prepare them to make that call.” Establishing working relationships and communications among
the diverse first responder agencies was suggested as well. One participant argued that: “I think
the leadership of municipal agencies and campus agencies is very important. They have to break
down barriers and build avenues by which the officers can become familiar with the geography of
the campus.”
5.2.2 Interactions between countermeasures and human behavior
In conjunction with identifying countermeasures for deterring active shooter incidents and
improving security, how these countermeasures impact the behavior of occupants and first
responder teams and vice versa were acknowledged as important issues during the interviews.
First, from the perspective of occupant response, while mass notification was commonly
mentioned as a countermeasure, concerns about its application were mentioned as well.
Participants underlined that the level of information conveyed by mass notification needs to be
very clear, and it should be communicated in such a way that occupants know how to respond to
the information. Otherwise, as claimed by a participant, “Mass notification could be a detriment
if appropriate response is not conveyed when occupants receive the notification, because then that
[not conveying appropriate information] just becomes chaos.” Additionally, participants
67

mentioned that access control could have a negative effect as well, because it sometimes acts as a
chokepoint, which could lead to large concentrations of people. It was also mentioned that not
having an area with concealment has benefits for active shooter incidents as well as general safety.
However, other participants expressed concerns about eliminating concealment or hiding places:
“Eliminating hiding places has a flipside as well, as the run-hide-fight program requires having
the ability to hide, so I think there should be a balance.” Another countermeasure that fell into this
category was related to line of sight. A participant commented that having open space could enable
first responder teams to find and stop the shooter more easily, but at the same time, shooters would
have line of sight as well to facilitate their attack. In contrast, the effectiveness of certain
countermeasures is also contingent upon occupant behavior. For mass notification, it was
suggested that occupants’ training and drills play significant roles. An example statement was:
“Occupants need to know how to respond appropriately to the advice that is being given [in mass
notification], and they need to know which exits or evacuation routes, or which zone of the building
they should be seeking refuge in.” Meanwhile, the participants stated that occupants’ natural
responses during active shooter incidents consist of seeking concealment, getting behind an object
or curling down under the desk and hide. Thus, the use of ballistic-resistant materials was identified
as being effective in this situation.
Second, from the perspective of first responder teams, one countermeasure identified that
could impact first responder teams was access control. A participant commented that: “If you
create a fortress-style building and a bad guy [a shooter] gets in, it would prevent or delay us
[first responder teams] from getting in, as we have to call the fire department to provide an access
to the building.” The effect of a remote locking control mechanism, as mentioned in the last
subsection, is also up to the authorities to use at their discretion in response to active shooter
68

incidents. Along this line, some participants highlighted countermeasures should be somewhat
tolerant to potential errors in their usage, as participants stated that: “Even if somebody [an
administrator] does not follow up on something [a certain procedure], such as forgetting to turn
on a switch or reviewing the camera footage, it [a countermeasure] still offers some level of
safeguard.” and “It is important to achieve the balance between providing the capability [for
administrators to take actions during active shooter incidents] and giving them the fewest number
of options to make sure they make the right decisions.”
5.2.3 Contextual influences on countermeasure implementations
The first context that was recognized to influence the implementation of countermeasures
was whether it is an internal or an external threat. Some participants argued that if it is an internal
threat, many of the countermeasures could be immediately circumvented, as shooters may not be
subject to the same access control and they would have familiarity with the environment, including
the location of occupants and rooms, and so on. Some countermeasures, however, were identified
as being immune to internal threats. An example was the remote locking control mechanism. A
participant commented that: “It [remote locking control mechanism] does not matter who you are,
if you are outside the door that is locked, you are not going to be able to get in.” Another option
that was mentioned included maintaining multiple egress routes, for which the inherited
assumption was that: “The shooter can only be in one place at a time, so if there are multiple
egress routes, they would provide evacuation opportunities for occupants, no matter who the
shooter is.” Moreover, the countermeasures for handling internal threats were discussed. It was
pointed out that there are software applications that manage electronic security systems, which
could be used to detect gunshots and deactivate certain badges in certain sections of the building.
Nevertheless, challenges are associated with this countermeasure as well, as a participant
69

commented that: “[To implement this countermeasure], you have to think about relocating
occupants out of the incident, as well as where the incident is occurring in relationship to other
sections of the building.” Similarly, to deal with internal threat, another countermeasure mentioned
was separating the main building with less secured areas, such as food delivery places, and
employing monitoring and alerting systems at these locations.
Building types were acknowledged as an influencing factor in countermeasure
implementations. It was stated that compared with school buildings, many office buildings consist
of a lobby or atrium areas. Thus, the participants mentioned that minimizing any “high ground”
vantage points (e.g., views from a mezzanine down to an atrium where events are held) was a
common countermeasure applied to office buildings. Moreover, it was pointed out that: “I think
in office buildings, …it is culturally more acceptable to have security desks for checking in and
turnstiles and access control for entering the building.” Regarding school buildings, it was
mentioned that high school design goes more towards open and flexible environments, instead of
being in the traditional classroom style. In addition, it was suggested that certain buildings, such
as houses of worship or very old buildings, might not even have countermeasures implemented
(e.g., access control), which imposes greater security risks.
During the interviews, the participants emphasized the significance of occupant
characteristics. For example, one participant explained: “Lower grade students have to follow the
instructions of their teachers, but high school students, probably middle school students as well,
are at an appropriate age for the run-hide-fight program.” Another participant analogized lower
grade students to patients in hospitals, because they must be supervised to a certain extent. In
addition, for the occupants that have limited mobility, such as those sitting in a wheelchair, the
participants suggested the use of safe rooms, so that: “Somebody who is disabled could get into
70

the safe room, lock it from the inside and be safe inside there, because running might not be a
feasible option.” Aside from occupants’ physical characteristics, the training that occupants
receive also has an impact. The participants mentioned that different schools, for example, have
their own training protocols and emphasis. Some may only conduct fire drills, some only teach
students “shelter in place”, and some focus more on the “run-hide-fight” program. The different
training that students receive, as the participants suggested, could largely impact occupant
behavior and the use of countermeasures.
5.2.4 Preparedness for different types of emergencies and normal operations
The participants identified that the cost associated with implementing countermeasures as
an important consideration. One participant argued that: “The question that should be asked is:
have the client, the school district or whoever owns the facility, done the analysis to see whether
the costs outweigh the risk?” However, another participant stated that “The problem with these
very low likelihood high consequence events is that nobody ever really thinks it will happen to
them to a degree, or that the risk reward ratio does not necessarily make sense in their minds.”
Along this line, a similar issue suggested by the participants was the cost associated with applying
certain countermeasures during an active shooter incident. For example, locking the entire building
down could lead to loss of revenue for a company, thus people might be afraid of using certain
countermeasures. It was further mentioned, “There needs to be a culture of not being afraid to
make the decision [of using a countermeasure].” On the other hand, while countermeasures cause
additional costs, one participant mentioned that parents are more likely to send their children to
the schools with appropriate security systems, which indicates that implementing countermeasures
could be beneficial in the economical aspect as well.
Besides the identified costs, other factors related to preparedness for active shooter
71

incidents and normal operation were aesthetics and people’s psychological perception of the
building design. One participant opined that buildings should at least appear to be welcoming and
inviting instead of looking like fortresses, and thus a balance between openness and security should
be taken into consideration. Moreover, people’s psychological reactions to the building design
were mentioned as well. An illustrative example was: “You might create an environment where
people do not feel safe to come to work, if there is a big sign that says: if an active shooter comes,
hide here.” Another operational factor that should be considered together with countermeasures
was recognized as occupants’ daily activities. One example mentioned by the participants was that:
“In our office building, we need to use badges to get in, however, everyone holds the door open
for anyone behind them, and this is a huge security risk.” Moreover, it was pointed out that
occupants might attempt to compromise certain countermeasures for the convenience of their daily
activities (e.g., propping a door open because it is too heavy).
Given that multiple types of emergencies can occur in buildings, the approaches to ensure
occupant safety are different. Thus, the interdependency of preparedness for different emergencies
was discussed during the interviews. For example, one participant claimed that: “Fire alarms are
good for fires, but they are not necessarily good for responding to an active shooter. Because fire
alarms would have occupants evacuate the building, going to a safe refuge area, which could be
a place where a lot of people gather, which could make it easier for an active shooter [to attack
occupants].” Other participants further illustrated this challenge by pointing out that even if there
are two separate security systems, shooters may still have a way to set off the fire alarm and have
everybody come out of their rooms. The complexity would further increase if there were multiple
types of emergencies (e.g., active shooter incidents, fires, earthquakes). A participant exemplified
that: “When the alarm goes off, I need to detect which type of alarms it is, and then I need to work
72

out how I should be respond[ing] to each type of the alarms”. Apart from occupant behavior,
countermeasures for different emergencies could also influence the building itself, as an example
statement being: “Ballistic resistant material is very heavy. In smaller buildings, adding that
weight to a building could impact how the structure performs under an earthquake.” That being
said, the participants also mentioned that there are similarities among different emergencies. For
example, knowing the location of building exits was considered important across different
emergency situations.
5.3 Discussion
In this study, we conducted focus group interviews to develop an understanding of current
countermeasures being used to proof buildings from active shooter incidents and protect building
occupants. The countermeasures mentioned during the interviews were primarily consistent with
the prevalent recommendations published by public agencies and research focus in prior studies
[118,133–135,203,204], which indicates that results of the focus group interviews reflect the state-
of-the-art of this area and have practical implications.
5.3.1 Influence of countermeasures on occupant behavior during active shooter incidents
The behavioral consequences (e.g., influence on human behavior) of certain
countermeasures were part of the main discussion during the interviews. In contrast with other
building emergencies such as fires and earthquakes, the duration of active shooter incidents is
typically very short. Between 2000 and 2013, 69.8% of the active shooter incidents ended within
five minutes [19]. Thus, before the arrival of first responder teams, occupants need to largely rely
on themselves to ensure their safety during building emergencies [8], hence more emphasis should
be placed on the influence of countermeasures on occupant behavior. In particular, occupant
behavior during building emergencies could be categorized into three phases: perception of the
73

environmental cues, decision-making and performing an action [205]. How current
countermeasures influence occupant behavior in these phases, based on the interview results, is
illustrated in Fig. 17.

Fig. 17. Influence of countermeasures on occupant behavior during active shooter incidents.
As shown in Fig. 17, it was noted that mass notification might impact occupants’
perception of environmental cues and decision-making during active shooter incidents either
positively or negatively, based on the information conveyed in the notification. This is in alignment
with previous research, which found that during emergency situations, providing accurate and
timely notifications would motivate occupants to take appropriate actions, whereas ambiguous
information might result in the opposite [50,134,206]. Meanwhile, participants mentioned that in
other emergency situations (e.g., fires and earthquakes), many occupants tend to ignore the
notification or alarms and continue pre-event behavior, because they may be perceived as false
alarms. This phenomenon is also reflected in the literature [104]. However, given the different
nature of emergency situations, whether occupants would still ignore notifications/alarms during
active shooter incidents is yet to be answered. Similarly, while mass notification provides audio
cues, visual access offers visual cues and could also influence occupants’ perception and decision-
making [207]. Compared with other emergencies (e.g., fires) where visual access mainly affects
74

occupants’ evacuation process [208], during active shooter incidents, visual access also influences
the behavior of first responder teams and shooters. Other countermeasures that affect occupants’
decision-making include access control, eliminating hiding places, etc., as these countermeasures
could change available options for occupants to respond. For example, access control was
acknowledged as a potentially negative factor because it limits the number of evacuation routes
and hence generates chokepoints at the main exit. Such effect was also shown in prior studies that
studied building emergency evacuations [27,209]. Therefore, it was underlined that maintaining
emergency evacuation exits at the same time could alleviate the negative effects of access control.
A significant factor, however, is that people would usually evacuate through the exits that they are
familiar with, such as the main exits [8]. Hence even though providing emergency exits could add
additional evacuation capacity, whether it will be sufficiently utilized remains uncertain and should
be assessed. Other countermeasures were considered to affect occupants’ action phase during
active shooter incidents by influencing their performance when carrying out a certain action as
well (e.g., the presence of multiple evacuation routes). As discussed above, countermeasures affect
occupant behavior from different perspectives (perceptual versus physical influences). Therefore,
to assess the correlation between occupant responses and countermeasures, a mixed approach
should be used. For those countermeasures that have physical impact (e.g., restricting the flow
rates), simulation-based methods could be used [27,88]; and those countermeasures that influence
occupant’s decision-making process, human-subject experiments could be more effective [55].
5.3.2 Training and drills for occupants and first responder teams
While the initial aim of this study was to assess building-related countermeasures, training
and drills were the top influencing factors that emerged during the interviews. It was emphasized
that countermeasures cannot perfectly deter active shooter incidents from happening or protect
75

occupants entirely. Thus, enhancing occupant and first responder teams’ preparedness to employ
the countermeasures and respond to the incidents appropriately is a decisive factor, which has also
been widely suggested in the literature for both occupants and first responder teams [210–212].
However, the participants also mentioned that in many locations across the U.S., building
occupants, especially students, are improperly trained. While it was mentioned that some high
schools have started to teach their students, faculty and staff the run-hide-fight program, the
participants pointed out that more students only practice fire drills and are not aware of the
response procedures recommended for active shooter incidents. It was suggested that the training
on responses, such as the run-hide-fight program, should trickle down to middle and primary
schools as well, but could be in different forms. For example, it was mentioned that for lower grade
students, the information conveyed during their training could be: “There might be a bad person
outside, and we want to hide from him or her,” instead of: “There may be somebody that is trying
to kill us.” Moreover, apart from conducting drills, other forms of training, such as serious games
that have been used in other emergencies [213,214], could be considered as well. More recently,
VR technology has been used to prepare occupants, teachers and administrators for active shooter
incidents [215]. More research could be done in this regard, to test the effectiveness of VR training
in this area. Apart from occupants’ usage of countermeasures during active shooter incidents, their
behavior during daily life also impacts the effectiveness of these countermeasures. For example,
occupants may (un)intentionally get around certain countermeasures (e.g., propping security doors
open for their convenience), which compromises the function of countermeasures. Thus, it is also
necessary to train occupants to increase their awareness of how to use countermeasures properly
in their daily life. In addition, training was found to be as important an issue for first responder
teams and building administrators, including both their responses to active shooter incidents as
76

well as coordination and communication among different agencies, which have also been reported
[133,216]. Gaining familiarity with the building was identified to be a crucial element for the
training of first responder teams as well. While the first responder teams may not have the
accessibility to on-site training in every building in a region or district, it was mentioned that
leveraging technologies, such as creating databases of building information on mobile devices that
first responder teams can access would be helpful for their operation.
5.3.3 Practical considerations for implementing countermeasures
While the primary goal of implementing countermeasures is to improve building security
and occupant safety in active shooter incidents, their influence on daily operations were also
stressed. Typical practical factors that impact the implementation of countermeasures include: (1)
cost efficiency (e.g., cost of implementing certain countermeasures and the influence on the
revenue of a company), (2) building aesthetics and attractiveness, and (3) risk level of different
types of attacks (e.g., external versus internal threats). Meanwhile, it was mentioned that providing
necessary information to occupants regarding the purpose of countermeasures is very important,
which highlights the importance of educating occupants via an appropriate approach when
implementing a countermeasure. A similar strategy has also been suggested in the literature [217].
In addition, different activities are correlated with building types (e.g., commercial activities for
office buildings and educational activities for school buildings). Previous studies have illustrated
that occupant responses are correlated with building functions as well as activities conducted in
buildings [5,73,135,218], hence a balance should be achieved between safety during active shooter
incidents and building functions.
5.3.4 Limitations and implications for future research
There exist several limitations in this study that could be further explored in future research.
77

First, the findings of this study, as with any other focus group studies, were based on the
participants’ pronouncements rather than defined facts, hence further studies are needed to justify
some of the results. In particular, one main finding from this study was the influence of
countermeasures on occupants, building administrators and first responder teams, which is directly
related to human behavior. Therefore, more empirical behavioral data could be collected from
human-subject experiments to verify the findings. Second, while we included participants with
expertise in building security, emergency management, and active shooter incidents in the focus
group interviews, some topics emerged during the interviews, such as cost efficiency of
countermeasures and building aesthetics, could be further explored from the perspectives of other
roles, such as building owners and occupants.
5.4 Conclusion
In response to the risk of active shooter incidents, a variety of security countermeasures
have been proposed and used in different types of buildings. While these countermeasures are
intended to proof buildings against active shooter incidents, challenges still exist in assessing the
effectiveness of countermeasures. In this study, we conducted three focus group interviews with
15 participants who have expertise in the area of building security and active shooter incidents.
We found that some of the countermeasures (e.g., mass notification, access control) can be a
double-edged sword, especially when it comes to their influence on the behavior of occupants and
first responder teams. Moreover, it was suggested that training and drills on how to use the
countermeasures and how to respond to active shooter incidents are of critical importance. Further
research in these areas is necessary. Practical factors that affect the use of countermeasures were
also discussed, including cost efficiency and occupants’ psychological perception of the building.
We also revealed that the use of countermeasures largely depends on the nature of the building
78

(e.g., different building types, different risk levels) and occupant characteristics (e.g., students of
different ages). Therefore, there is no universal solution that works for the whole spectrum of
environments, and thorough investigations are needed to develop countermeasures for active
shooter incidents that cater to different environments. Research questions 2.1 – 2.4 were addressed
in this chapter.

79

Chapter 6. The Influence of Security Countermeasures on Human Behavior
during Active Shooter Incidents
As demonstrated in Chapter 5, a critical consideration of countermeasure implementation
for active shooter incidents is the influence on human behavior. Certain countermeasures may be
beneficial for building security but limit people’s response options during active shooter incidents.
Nevertheless, empirical assessment of countermeasures with regards to their behavioral influence
is lacking in the literature. In this chapter, we present a study that evaluates the influence of
countermeasures on human behavior during active shooter incidents in office and school buildings.
The remainder of this chapter is organized as follows. Section 6.1 describes the methodology of
this study. Section 6.2 presents the experiment results, including ecological validity, participants’
response time and decisions, as well as participants’ subjective responses during the experiment.
Section 6.3 discusses implications of the findings and limitations in this study. Finally, section 6.4
concludes the chapter.
6.1 Methodology
6.1.1 Virtual environment
An office building and a school building were geometrically modeled to simulate an active
shooter incident in indoor environments, as these two buildings have been most frequently targeted
by shooters [2]. The two buildings were both based on a Revit model provided by an architecture,
design, planning and consulting firm. A licensed architect further reviewed the models to ensure
that they were representative of real-world office and school buildings. The first-story layouts of
the two buildings are shown in Fig. 18. Both buildings had 7 exits, 5 staircases, and a cafeteria and
a kitchen exactly in the same locations. To avoid any potential confounding factors, building size
and structure were kept identical between the two buildings. Several manipulations were adopted
80

merely to represent the context of an office and a school without changing the building layout or
complexity: in the office, a conference room replaced the teacher’s lounge, an office room replaced
the medical room, and office cubicles replaced classrooms in the school.

Fig. 18. First-floor layouts of the standard office (left) and school (right). Stars denote
participants’ starting locations: one in the cafeteria and one in the hallway. Arrowed lines denote
the shooter’s movement trajectory.
To assess the impact of security countermeasures on participants’ responses to the active
shooter incident, a list of countermeasures was identified by conducting a comprehensive literature
review and focus group interviews with 15 building designers/engineers, security experts, and law
enforcement professionals [219,220]. The countermeasures were implemented and resulted in
another version of the office and school (hereafter referred to as enhanced buildings as opposed to
standard buildings, which had no countermeasures). The description and implementation of the
countermeasures are presented in Table 7. In total, four different virtual environments were
included in the experiment, namely (1) standard school, (2) standard office, (3) enhanced school,
and (4) enhanced office.

81

Table 7. Security countermeasures included in the virtual environments.
Countermeasures Intention
Comparison of standard and enhanced buildings
Standard Enhanced
Eliminating
hiding places
Preventing shooters
from hiding

Installing barriers Preventing shooters
from entering the
building

Isolating
unsecured areas
from secured
areas
Preventing shooters
from entering the
building and providing
buffer zones

Access control Preventing shooters
from entering the
building

Using frosted
windows
Limiting shooters’ sight
and providing
concealment for
occupants

Staggering
interior doors
(school)
Limiting shooters’ sight
and the effect of
bullets/blast through the
building

Staggering
interior doors
(office)
Limiting shooters’ sight
and the effect of
bullets/blast through the
building

Non-player characters (NPCs) as building occupants were included in the virtual
82

environments to increase realism and represent the natural effect that crowd behavior has during
emergency situations. Each virtual environment contained 81 building occupants, which was
determined considering the tradeoffs between the level of realism and computational burden. 25
occupants were initially in the cafeteria, 12 were in the hallways, 4 were in the teacher’s
lounge/conference room, and the remaining 40 were in the office cubicles/classrooms. It is
noteworthy that while we kept the occupants’ appearance, gender, and clothing the same across
different conditions to avoid potential confound, we intentionally modeled them to look relatively
young so that they can reflect the corresponding social contexts (office workers in the office and
students in the school). Fig. 19 (a) and (b) illustrate some occupants at their initial locations. The
occupants were given different response times ranging from 0 to 10 seconds when the shooting
started, which were determined based on the occupants’ distance to the shooter and were reviewed
by a former FBI agent to ensure realism. After occupants started to respond to the shooting, they
would navigate to preprogrammed destinations. We distributed the destinations among different
exits and hiding places to reflect the natural difference between the standard and enhanced
buildings and avoid biasing participants’ decisions. For example, if barriers were implemented in
the outdoor dining area, naturally occupants could not evacuate there, and instead they were
assigned to other exits according to their initial locations. Other than these, occupants’ behaviors
remained the same among different conditions.

(a) (b) (c)
Fig. 19. Virtual occupants and the virtual shooter. (a) occupants from the view of the starting
location in the cafeteria. (b) occupants from the view of the starting location in the hallway. (c)
83

shooter entering the building from the main entrance.
The shooter was represented by a male character (Fig. 19 (c)), as most active shooter
incidents in the U.S. involved a single shooter and the majority of them were males [2]. The
shooting started 10 seconds after the beginning of the experiment and ended after 105 seconds.
The shooter would first enter the building from the main entrance and go to the cafeteria. Then he
would move along the hallway and leave the building via Exit 3. During the incident, the shooter
would shoot the occupants that are visible to him. Once being shot, occupants would fall onto the
ground and stay still, however, no blood would appear to avoid violent and graphic content. It is
important to note that at the beginning of the experiment, participants were not in close proximity
to the shooter and could not directly see the shooter. Nevertheless, participants may encounter the
shooter during the experiment depending on the directions they decide to go. While this
information was not provided to the participants, the shooter could not shoot them during the
experiment so that all participants could complete their trials.
6.1.2 Apparatus and simulations
The virtual environments were initially created for in-person experiments using VR
headsets and controllers. However, due to the COVID-19 pandemic, we had to modify the virtual
environments as WebGL applications and deployed them on GitHub. Such a setup allowed
participants to access the virtual environments remotely in a web browser using laptops/desktops,
keyboards, and mice for social distancing. Unity game engine was used to develop the virtual
environments. Sound effects, including occupants’ chatting and panic sound and shooting sound
(represented a semi-automated assault rifle, which was recommended, reviewed, and approved by
a former FBI agent) were added to the virtual environments. Moreover, participants’ locations in
the virtual environments and the corresponding timestamps were updated and recorded twice every
second during the experiment.
84

6.1.3 Experiment design
This study used a 2 (design: standard or enhanced) × 2 (order: office before school or
school before office) × 2 (building: office or school) mixed-subjects designs, where the latter factor
was within-subjects. Starting location (cafeteria or hallway) for the first trial was also randomized,
and the other location was then used for the second trial. As a result, there were 8 experiment
groups, as shown in Table 8. Each group consisted of two trials and each participant was randomly
assigned to one of the groups. Between the two trials, building design was kept the same, whereas
building types, order, and starting location were different. The metrics to measure participants’
responses included their response time and decision. The response time was defined as the time
participants spent from the start of the experiment to the moment that they completed the last action
(e.g., evacuating the building via an exit or hiding at a place). Decision referred to whether
participants chose to run, hide, or fight, and what specific destination (e.g., an exit or hiding place)
they chose. Response time and decision were determined by reviewing video recordings of the
experiment and the recorded data. Moreover, to examine participants’ perceptions and
considerations during the experiment, subjective questions were asked in the surveys.
Table 8. Experiment groups.
Group Design First trial Second trial
1 Standard Office, cafeteria School, hallway
2 School, hallway Office, cafeteria
3 Office, hallway School, cafeteria
4 School, cafeteria Office, hallway
5 Enhanced Office, cafeteria School, hallway
6 School, hallway Office, cafeteria
7 Office, hallway School, cafeteria
8 School, cafeteria Office, hallway
6.1.4 Participants
A total of 162 adults (36.73 ± 11.14 years-old on average, 71 males and 91 females, 79
middle/high school teachers and 83 office workers) participated in the study. 5 participants were
85

excluded due to technical issues with the virtual environments, or the participants did not finish
the experiment. All participants received a $20 Amazon Gift Card as monetary incentives after the
study. Participants were recruited nationwide via social media platforms (e.g., Facebook, Craigslist)
and emails, and their occupations were verified by asking them to provide a work email address
or an ID. All participants reviewed the informed consent and agreed to participate in the
experiment. This study was approved by the University Park Institutional Review Board (UPIRB)
of University of Southern California. Since minors are considered to be more vulnerable or at-risk,
we only recruited adults in this study to limit risks. All experiments were performed in accordance
with relevant guidelines and regulations.
6.1.5 Procedure
A Zoom meeting was used to connect participants with the experimenter. At the beginning,
participants reviewed an information sheet, which described that the objective of this study was to
examine how building designs influence human responses during emergency scenarios. Following
the informed consent, participants completed a training session, in which they followed the
experimenter’s instructions to get familiar with different actions (i.e., turn, walk, run, and crouch)
in the virtual environments. The shooter was not included in the training environment in order not
to reveal the type of emergency they would experience in the experiment and participants could
not directly fight against the shooter to ensure all participants experienced the same shooting
scenario during the same amount of time. That being said, participants were not informed that they
could not fight either, hence some participants may still choose to confront the shooter.
Accordingly, the “fight” decision in this study denotes those participants confronted or approached
the shooter. The training took place in a virtual outdoor space, which was different from the
experiment environment. Next, participants completed a pre-experiment survey, which asked their
86

demographic information and positive and negative emotions measured with the Positive Affect
and Negative Affect Scale (PANAS) from 1 (not at all) to 5 (extremely) [170].
Participants then familiarized themselves with the building by following a virtual tour
guide, as shown in Fig. 20. The virtual guide was preprogrammed to follow a given trajectory to
show participants the different sections of the building. The tour took approximately 4 minutes.
The participants were asked to familiarize themselves with the building without pointing out
countermeasures implemented in the enhanced building to avoid potential confounding effects.
Subsequently, participants responded to the active shooter incident in the same building (i.e., first
trial). Upon the completion of the first trial, participants responded to a mid-experiment survey,
which asked their perceptions of the shooter (from 1 as “neutral” to 5 as “worst thing ever”) and
occupants (from 1 as “extremely negative” to 5 as “extremely positive”), whether they considered
different factors (shooter and occupant behavior, exits, hiding places, and occupant safety) when
responding to the active shooter incident from 1 (strongly disagree) to 5 (strongly agree), and
positive and negative emotions measured with PANAS. Specifically, with regards to the
perception of the shooter and occupants, we asked participants to rate their high-level perceptions
about the presence, appearance, and behavior of the shooter and occupants.

Fig. 20. Virtual tour guide for the building tour.
6.1.6 Analysis
87

Repeated measures analysis of covariance (ANCOVA) was mainly used for the analysis.
Since each participant experienced two trials with different building types and starting locations
(see Experiment design), building type was entered as the within-subjects factor and starting
location was entered as the covariate. Between-subjects factors included design, occupation, and
order. Dependent variables in the repeated measures ANCOVA included participants’ response
time, decisions, and subjective perceptions. To analyze decisions, we considered participants’
choices among run, hide, and fight. Because the order recommended by authorities for surviving
an active shooter incident is run, then hide, then fight [41], a continuous variable was created where
higher scores are further down this recommended order of actions (1: run, 2: hide, 3: fight).
Moreover, to reveal if any individual difference variables affected participants’ response time and
decisions, we conducted linear regression analysis with participants’ age, gender, and previous
experience with active shooter incidents as independent variables. The significance level was set
at 0.05 and the marginal significance level was set as 0.10. All the data analysis was conducted
using the SPSS 25 software.
6.2 Results
6.2.1 Ecological validity
Before analyzing participants’ responses to the active shooter incident, we examined their
emotional responses and sense of presence to assess the ecological validity of our virtual
environments. For emotional response, a 2 (valence: positive or negative) × 3 (time: before the
first trial, after the first trial, or after the second trial) × 2 (design: standard or enhanced) × 2 (order:
experienced the office before school or experienced the school before office) × 2 (occupation:
office workers or teachers) factorial ANCOVA was conducted, with valence and time as within-
subjects factors, design, order, and occupation as between-subjects factors, and location (started
88

in the hallway or kitchen) as the covariate. The results revealed a significant valence × time (𝐹 2,296

= 8.346, 𝑝 = 0.001, 𝜂 𝑝 2
= 0.053) interaction, such that participants’ positive emotions decreased
consistently during the two trials, and negative emotions increased significantly during the first
trial and decreased slightly during the second trial (Fig. 21). This result is in accordance with our
previous findings [221], which showed that the virtual environments evoked a large increase of
participants’ negative emotions. Given the complex nature of our experiment design, there were
also inscrutable valance × time × design (𝐹 2,296
= 9.319, 𝑝 < 0.001, 𝜂 𝑝 2
= 0.059) and valence ×
time × occupation (𝐹 2,296
= 3.018, 𝑝 = 0.062, 𝜂 𝑝 2
= 0.02) interactions. There were also other main
effects and lower-order interactions, which were less relevant as they collapsed across
time/valence or were qualified by the above effects.

Fig. 21. The effect of valence × time interaction on participants’ emotional arousals.
We asked participants’ sense of presence only once (asking them to estimate across both
trials) in the post-experiment survey, thus we conducted a 2 (design) × 2 (order) × 2 (occupation)
between-subjects factorial ANCOVA, with location as the covariate. There was a marginally
significant main effect of order (𝐹 1,148
= 3.036, 𝑝 = 0.084, 𝜂 𝑝 2
= 0.02), such that the participants
experiencing the office before school (M = 135.73) had a lower sense of presence than those
experiencing the school before office (M = 142.05), perhaps because a school shooting is more
psychologically engaging. More importantly for our validation, sense of presence averaged 4.63
89

across all participants on a scale from 1 to 7. The results indicated that our virtual environments
induced significant emotional arousals and adequate sense of presence for the participants, which
supports validity of the following results.
6.2.2 Response time and decision
We first analyzed whether participants’ individual differences affected their response time,
hence a linear regression analysis was conducted with response time as the dependent variable and
participants’ age, gender, and previous experience with active shooter incidents as independent
variables. The results showed that response time in the office was significantly affected by
participants’ age (β = 0.192, 𝑝 = 0.018), such that younger participants had a shorter response time
than older participants. Next, to evaluate how participants’ response time was affected by the
experimental conditions, we conducted a 2 (building: office or school) × 2 (design) × 2 (order) ×
2 (occupation) factorial ANCOVA, with building as the within-subjects factor, design, order, and
occupation as between-subjects factors, and location as the covariate. The results revealed a
marginally significant main effect of building (𝐹 1,148
= 3.437, 𝑝 = 0.066, 𝜂 𝑝 2
= 0.023), such that
participants spent less time in the office (M = 38.95 seconds) than school (M = 43.73 seconds).
This effect was qualified by an interaction with order (building × order, 𝐹 1,148
= 11.837, 𝑝 =
0.001, 𝜂 𝑝 2
= 0.074): participants had a shorter response time in the office if they experienced the
school before office (M = 36.66 seconds for the office, M = 49.07 seconds for the school).
Moreover, although it was a smaller difference, they had a shorter response time in the school if
they experienced the office before school (M = 41.23 seconds for the office, M = 38.39 seconds
for the school). As building and order relate to participants’ first and second trials, this result
demonstrated that participants spent less time responding to the shooting in the second trial, and
this was especially true when the office was second. Likewise, a marginally significant main effect
90

of design (𝐹 1,148
= 2.799, 𝑝 = 0.096, 𝜂 𝑝 2
= 0.019) was found, such that participants in standard
buildings (M = 38.84 seconds) had a shorter response time compared with those in enhanced
buildings (M = 43.84 seconds). This result indicated that, with the implementation of
countermeasures in enhanced buildings, participants might have needed more time to reach a
desired destination. Given the complex nature of our experiment design, there was also an
inscrutable design × order × occupation (𝐹 1,148
= 3.149, 𝑝 = 0.078, 𝜂 𝑝 2
= 0.021) interaction
1
.
To analyze participants’ choices among run, hide, and fight, we first examined the impact
of individual differences by conducting the same linear regression analysis as for response time.
The results showed that participants’ decisions in the office were significantly affected by their
previous experience with active shooter incidents (β = 0.167, 𝑝 = 0.039), such that in the office,
participants with previous experience with active shooter incidents were more likely to hide in
place. Next, we conducted another ANCOVA for participants’ decisions with building as the
within-subjects factor, design, order, and occupation as between-subjects factors, and location as
the covariate. This analysis revealed a significant building × order × occupation (𝐹 1,148
= 4.206,
𝑝 = 0.042, 𝜂 𝑝 2
= 0.028) interaction. Since there were a total of only five participants chose to fight
(4 office workers and 1 teacher), the interaction effect revealed that office workers scored lower
and thus chose “run” more often in the second trial than the first, whereas teachers scored higher
in the second trial and thus chose to “hide” more in the second trial than the first, as shown in Fig.
22A. This result indicated that, with repeated trials, people would fall more into the expected or
normative patterns associated with their occupation (e.g., teacher to stay to protect students). Main

1
There was an inscrutable design × order × occupation interaction (𝐹 1,148
= 3.149, 𝑝 = 0.078, 𝜂 𝑝 2
= 0.021) on participants’ response
time: the office workers experiencing the office before school spent significantly less time in standard buildings (M = 29.78) than
enhanced buildings (M = 47.26). However, the impact of design on response time was not as significant for teachers or office
workers experiencing the school before office.

91

effects of design ( 𝐹 1,148
= 23.718, 𝑝 < 0.001, 𝜂 𝑝 2
= 0.138) and order ( 𝐹 1,148
= 3.058, 𝑝 =
0.082, 𝜂 𝑝 2
= 0.02) were found and qualified by a marginally significant design × order (𝐹 1,148
=
2.733, 𝑝 = 0.098, 𝜂 𝑝 2
= 0.018) interaction: participants were more likely to run instead of hiding
in standard buildings than in enhanced buildings, especially for those experiencing the office
before school (Fig. 22B). This result suggested that the implementation of certain countermeasures
(e.g., access control) might have impeded participants’ ability to run in enhanced buildings.

Fig. 22. The effects of building × order × occupation interaction and design × order interaction
on participants’ choices.
Next, we examined participants’ decisions at a more fine-grained level by looking into their
choices of specific exits or hiding places. As no within-subjects effect was found for participants’
decisions of run, hide, or fight in the above analysis, the two trials were collapsed here. Specifically,
we created another continuous variable to denote whether a choice was made 0, 1, or 2 times by a
participant across the two trials. For each exit and hiding place, a 2 (design) × 2 (order) × 2
(occupation) between-subjects factorial ANCOVA was conducted, with location as the covariate.
The results are shown in Table 9. First, as Exits 4, 5, 6 and 7 were removed in enhanced buildings,
we investigated participants’ choice of evacuating via Exit 3 (at the end of the hallway). The result
revealed a significant main effect of design, such that participants in enhanced buildings (M = 0.61)
evacuated via Exit 3 more frequently than those in standard buildings (M = 0.18). Second, there
92

were main effects of design and order on participants’ choice of hiding in the kitchen, such that
participants in enhanced buildings (M = 0.25) or those experiencing the school before office (M =
0.2) chose to hide more in the kitchen than those in standard buildings (M = 0.03) or experiencing
the office before school (M = 0.07). Similarly, in enhanced buildings, participants chose to hide
more often under staircases (M = 0.17 for standard buildings, M = 0 for enhanced buildings) or in
the conference room/teacher’s lounge (M = 0.04 for standard buildings, M = 0.14 for enhanced
buildings). The above results suggest that the implementation of countermeasures in enhanced
buildings shaped participants’ choices of specific exits and hiding places, and the order of
experiencing the office and school also played a role in participants’ choices.
Table 9. Effects of design and order on participants’ choices.
Dependent measure Effects F P - value 𝜂 𝑝 2

Evacuating via Exit 3 Design 20.005 < 0.001 0.119
Hiding in the kitchen Design 16.409 < 0.001 0.1
Order 5.272 0.023 0.034
Hiding under staircases Design 15.261 < 0.001 0.093
Hiding in the conference room/teacher’s lounge Design 4.713 0.032 0.031
6.2.3 Subjective responses
We also examined participants’ subjective responses to the active shooter incident in terms
of perceptions of the shooter and other occupants, reports of how much their responses were
influenced by the shooter and other occupants, and finally, the degree to which participants
reported considering other factors (i.e., exits, hiding places, occupant safety) in making their
decisions. For each subjective response, we again conducted a 2 (building) × 2 (design) × 2 (order)
× 2 (occupation) factorial ANCOVA, with building as the within-subjects factor, design, order,
and occupation as between-subjects factors, and location as the covariate.
First, we analyzed participants’ perceptions of the shooter and other occupants. Main
effects of occupation (𝐹 1,148
= 4.665, 𝑝 = 0.032, 𝜂 𝑝 2
= 0.031) and order (𝐹 1,148
= 3.103, 𝑝 =
93

0.08, 𝜂 𝑝 2
= 0.021) were found for participants’ perceptions of the shooter: office workers (M = 4.51)
and participants experiencing the school before office (M = 4.48) perceived the shooter more
negatively, compared with teachers (M = 4.2) and participants experiencing the office before
school (M = 4.23), who might have possibly had more exposure to (because of traits associated
with their occupation) or less judgement of (because office is not as cruel of a target as school) the
shooter. Considering perceptions of other occupants, a marginal main effect of building (𝐹 1,148
=
3.564, 𝑝 = 0.061, 𝜂 𝑝 2
= 0.024) was found, but it was qualified by a significant building ×
occupation (𝐹 1,148
= 5.852, 𝑝 = 0.017, 𝜂 𝑝 2
= 0.038) interaction: office workers had more positive
perceptions of other occupants in the office, whereas teachers had more positive perceptions of
other occupants in the school (Fig. 23). Even though the occupants had the same appearance in the
office and school, participants liked the occupants better when they were in their own occupational
context.

Fig. 23. The effect of building × occupation interaction on participants’ perceptions of other
occupants.
We next evaluated participants’ reports of how their responses were influenced by the
shooter and other occupants. For influence the shooter had on responses, given the complex nature
of our experiment design, inscrutable interactions
2
were found for design × order × occupation

2
There were inscrutable design × order × occupation (𝐹 1,148
= 3.221, 𝑝 = 0.075, 𝜂 𝑝 2
= 0.021) and building × design × order ×
occupation (𝐹 1,148
= 3.52, 𝑝 = 0.063, 𝜂 𝑝 2
= 0.023) interactions on influence the shooter had on participants’ response: in the first
94

(𝐹 1,148
= 3.221, 𝑝 = 0.075, 𝜂 𝑝 2
= 0.021) and building × design × order × occupation (𝐹 1,148
= 3.52,
𝑝 = 0.063, 𝜂 𝑝 2
= 0.023). Considering influence of other occupants, a significant building × order
× occupation (𝐹 1,148
= 5.673, 𝑝 = 0.019, 𝜂 𝑝 2
= 0.037) interaction was found, such that office
workers were less influenced by other occupants in the second trial, whereas teachers were more
influenced by other occupants in the second trial (Fig. 24A). This result indicated that in the second
trial, office workers possibly thought more about self-preservation, whereas teachers considered
more about the safety of others. There was also a significant design × occupation (𝐹 1,148
= 4.775,
𝑝 = 0.03, 𝜂 𝑝 2
= 0.031) interaction, such that office workers were more influenced by other
occupants in the standard buildings, whereas teachers were more influenced by other occupants in
enhanced buildings (Fig. 24B). It is possible that office workers might focus more on others in the
standard building because they chose to follow other occupants when evacuating the building.
However, when running was less of an option in enhanced buildings, they did not pay as much
attention to other occupants’ behavior. Given the complex nature of our experiment design, there
was also an inscrutable design × order × occupation (𝐹 1,148
= 3.519, 𝑝 = 0.063, 𝜂 𝑝 2
= 0.023)
interaction
3
.

trial, office workers (M = 4.53 for standard buildings, M = 4.01 for enhanced buildings) were more influenced by the shooter in
standard buildings, whereas teachers (M = 4.17 for standard buildings, M = 4.57 for enhanced buildings) were more influenced by
the shooter in enhanced buildings. In the second trial, shooter influence was similar for both office workers (M = 4.48 for standard
buildings, M = 4.37 for enhanced buildings) and teachers (M = 4.36 for standard buildings, M = 4.42 for enhanced buildings) in
standard and enhanced buildings.

3
There was an inscrutable design × order × occupation interaction (𝐹 1,148
= 3.519, 𝑝 = 0.063, 𝜂 𝑝 2
= 0.023) on influence other
occupants had on participants’ response: for those experiencing the school before office, office workers (M = 4 for standard
buildings, M = 3.25 for enhanced buildings) were more influenced by occupants in standard buildings, whereas teachers (M = 3.5
for standard buildings, M = 4.03 for enhanced buildings) were more influenced by occupants in enhanced buildings.
95

Fig. 24. The effects of building × order × occupation interaction and design × occupation
interaction on participants’ ratings of occupant influence.
The degree to which participants reported considering other factors (i.e., exits, hiding
places, occupant safety) when making their decisions was also analyzed. First, there was a
significant building × order ( 𝐹 1,148
= 8.639, 𝑝 = 0.004, 𝜂 𝑝 2
= 0.055) interaction on the
consideration of exits, such that participants considered exits more in the second trial (M = 4.45)
than the first (M = 4.11). This result suggested that participants might have thought more about
how not to get trapped in the building in the second trial, after having already experienced the
shooting. Given the complex nature of our experiment design, there was also an inscrutable
building × design × order × occupation (𝐹 1,148
= 4.089, 𝑝 = 0.045, 𝜂 𝑝 2
= 0.027) interaction
4
.
Second, there was a significant building × occupation (𝐹 1,148
= 10.54, 𝑝 = 0.001, 𝜂 𝑝 2
= 0.066)
interaction on the consideration of hiding places, such that office workers considered hiding places
more in the office, whereas teachers considered hiding places more in the school (Fig. 25). This
result suggested that participants were more inclined to stay and hide in a building that matched
their occupational context, perhaps because of higher familiarity with the building. Third, even
though participants did not have direct interactions (e.g., talking, non-verbal coordination) with

4
There was an inscrutable building × design × order × occupation interaction (𝐹 1,148
= 4.089, 𝑝 = 0.045, 𝜂 𝑝 2
= 0.027) on
participants’ consideration of exits: office workers (M = 4.03 for the first trial, M = 4.49 for the second trial) tended to be more
influenced by exits in the second trial, whereas teachers experiencing standard buildings (M = 4.34 for the first trial, M = 4.29 for
the second trial) were influenced by exits approximately the same between the two trials.
96

other occupants, we found a significant main effect of occupation (𝐹 1,148
= 11.192, 𝑝 = 0.001, 𝜂 𝑝 2

= 0.07) on participants’ concerns for the safety of other occupants: teachers (M = 3.04) had more
concerns for other occupants’ safety compared with office workers (M = 2.42), possibly because
their occupation requires supervision and caregiving for others, where office work does not. Again,
given the complex nature of our experiment design, there was also an inscrutable design × order ×
occupation (𝐹 1,148
= 5.294, 𝑝 = 0.023, 𝜂 𝑝 2
= 0.035) interaction
5
.

Fig. 25. The effect of building × occupation interaction on participants’ consideration of hiding
places.
6.3 Discussion
The results of this study revealed that building design played an important role in
participants’ responses to active shooter incidents. In enhanced buildings where countermeasures
were implemented, participants had a longer response time and fewer decisions of “run.” Indeed,
since access control was implemented in enhanced buildings, several exits became inaccessible.
Participants who decided to evacuate had to travel a longer distance in enhanced buildings, such
as evacuating via Exit 3 at the end of the hallway. In standard buildings, however, the availability

5
There was an inscrutable design × order × occupation interaction (𝐹 1,148
= 5.294, 𝑝 = 0.023, 𝜂 𝑝 2
= 0.035) on participants’
concerns for the safety of other occupants: office workers experiencing the office before school (M = 1.95 for standard buildings,
M = 2.72 for enhanced buildings) had more concerns for occupant safety in enhanced buildings, which also applied for teachers
experiencing the school before office (M = 2.88 for standard buildings, M = 3.4 for enhanced buildings). In contrast, if office
workers experienced the school before office (M = 2.66 for standard buildings, M = 2.32 for enhanced buildings) or teachers
experienced the office before school (M = 2.98 for standard buildings, M = 2.89 for enhanced buildings), they had more concern
for occupant safety in standard buildings.

97

of more exits allowed participants to choose a closer exit. This finding was consistent with several
previous studies, which demonstrated that people prioritize to choose exits in close proximity
during evacuation, as their primary objective is to stay away from emergencies [16,222]. While
access control is important for building security, eliminating entry (or exit) points to a building
may be detrimental to human safety during active shooter incidents.
Participants in the enhanced buildings also chose to hide significantly more in the kitchen
and conference room/teacher’s lounge, possibly due to the decreased ability to evacuate. More
hiding decisions in enhanced buildings might also be related to the implementation of frosted
windows in the kitchen and conference room/teacher’s lounge, as staying out of shooters’ sight is
crucial for survival and is frequently mentioned in emergency trainings [41,223]. In contrast, even
though some participants chose to hide under staircases in standard buildings, they were incapable
of doing so in enhanced buildings due to the removal of storage rooms to prohibit shooters to hide.
This result suggested that hiding is an essential component of people’s responses to active shooter
incidents. Yet, hiding behavior has been underexplored, with only a few studies that evaluated
hiding behavior using predefined rules (e.g., choose to hide based on the distance to shooters)
[120,131]. People should not be assumed to have a sole response to active shooter incidents, and
a balance should be achieved between improving building security and providing multiple
response options. This result is consistent with a recent study, which demonstrated that a multi-
option response plan was more effective than the traditional lockdown approach [224]. Overall,
our study revealed that human-building interactions should be taken into consideration to improve
human safety, as echoed by several recent studies [47,225,226].
In this study, building types were different between the two trials. The results revealed that
both building types and trial order affected participants’ responses: participants in the school spent
98

a longer time than those in the office. A possible reason for this finding is that participants might
be generally more familiar with offices, which facilitated their reading of the building typology
and perception of emergency cues [123,227]. Moreover, although unexpected, we found the order
of experiencing the office and school had a significant effect on participants’ responses across the
two trials. The participants experiencing the office before school had more decisions of “run” in
standard buildings, compared with those experiencing the school before office. Similarly, the
participants experiencing the school before office had more negative perceptions of the shooter.
Such effect of order might be attributed to the consistency of participants’ behavior/perception
when experiencing similar scenarios or making multiple decisions [228]. When experiencing the
standard office in the first trial, participants’ first intuition might be to evacuate immediately rather
than consider the safety of others. Similarly, when experiencing the school in the first trial,
participants might perceive the shooter more negatively as students are the victims. Depending on
the building context in the first trial, participants’ behavior/perception was carried over to the
second trial. Additionally, our results showed that participants had a shorter response time in the
second trial. To interpret this finding, we investigated participants’ subjective responses after the
second trial and unsurprisingly, we found participants had very high familiarity with the building
(M = 4.28 on a scale from 1 to 5) and emergency (M = 4.127). This result indicated that greater
familiarity might allow participants to respond more efficiently, which is consistent with previous
findings [8]. Such impact of familiarity also suggest that training is an effective approach to
improve human safety, as people can be informed about appropriate responses and obtain
increased familiarity with different emergency scenarios [229]. Future research could investigate
the effectiveness of training for people’s responses during active shooter incidents, especially with
an emphasis on innovative training approaches to improve training effectiveness [219].
99

With regards to participants’ roles (occupational backgrounds), several interesting findings
were identified. First, teachers had significantly more concerns for occupant safety than office
workers. Teachers are routinely trained to care for students when an emergency occurs, which
might have a direct impact on their responses to active shooter incidents [215]. More interestingly,
the responses of office workers and teachers were correlated with the building and social contexts.
For instance, even though the virtual occupants had the same appearance in the office and school,
office workers had more positive perceptions of occupants in the office, whereas teachers had more
positive perceptions of occupants in the school. This finding had to do with participants’
psychological ties with others, as they tended to have closer relationships with familiar people or
those sharing common identities [34,62]. Additionally, office workers considered hiding places
more in the office, while teachers considered hiding places more in the school. This finding could
be attributed to participants’ confidence in the building context: people are usually more confident
to shelter in place in a building they are frequently exposed to, whereas if they are in an unfamiliar
building, leaving the scene tends to be their first resort [230]. Finally, our results revealed that in
contrast to office workers, teachers had more “hide” decisions and were more influenced by other
occupants in the second trial. This finding again suggested the impact of participants’ daily roles.
For office workers, the main goal is self-preservation. For teachers, on the other hand, they need
to consider a duty to care for other occupants in addition to themselves. This is further reflected in
participants’ feedback during the experiment: some teachers mentioned that during the first trial,
they acted reactively due to the sudden onset of shooting, whereas in the second trial, they
considered more the responsibility of being a teacher.
Overall, the answers to our three research questions were all “yes.” First, some building
design-related countermeasures aimed for improving building security could affect participants’
100

response time and choices among run, hide, and fight. Second, participants had a longer response
time in the school than office and their response time was significantly reduced in the second trial.
Third, office workers had more positive perceptions of other occupants in the office, whereas
teachers had more positive perceptions of other occupants in the school. Teachers also had more
concerns for other occupants’ safety. Our research findings showed that there is a tradeoff between
preventing the occurrence of active shooter incidents and improving occupant safety when an
incident occurs. Moreover, the effectiveness of countermeasures can vary in different shooting
scenarios, hence appropriate strategies should be adopted given the specific context [135]. For
instance, the primary population and function of a building, as well as people’s familiarity with
the building should be taken into consideration when developing and implementing
countermeasures.
Moreover, our results demonstrated the effectiveness of virtual environments for
behavioral studies, especially in situations that would not be safe or feasible to study in person
(e.g., building emergencies, proximal threats) [231,232]. That being said, there were still
limitations associated with this study. First, while our findings show that people’s roles affect their
responses to active shooter incidents, we only recruited office workers and teachers in this study,
hence the results may not be generalizable to other population groups. Also, even though we
recruited participants nationwide, the results may not apply to all office workers and teachers as
other factors, such as their educational levels and previous training experience can affect their
behavior as well. Second, while multiple countermeasures were examined in this study, some
building attributes (e.g., width of exits and hallways) that depend on macroscopic characteristics
of human behavior (e.g., crowd flow rate) cannot be easily examined using virtual experiments.
Crowd simulations could be a more appropriate approach for examining these countermeasures,
101

which should be further explored by future research [51]. Third, a countermeasure could be
implemented in multiple ways. For example, barriers, vehicles, physical guards, and limiting
entrance/exits could all be used for the purpose of access control. In this study, the implementation
of countermeasures was based on the results of our literature review and focus group interviews
[219,220]. Other forms of implementation could be explored and examined in future research.
6.4 Conclusion
Humans, buildings, and emergencies are three interconnected components. Their
relationships should be considered altogether for improving human safety during building
emergencies. However, human behavior has not been sufficiently considered when developing and
implementing countermeasures for active shooter incidents. In this study, we conducted virtual
experiments to empirically examine the impact of countermeasures on people’s responses to active
shooter incidents, with participants’ occupation as well as building and social contexts taken into
consideration. Our findings highlighted that some countermeasures intended to improve building
security may negatively affect people’s response to active shooter incidents. The emergency
context and daily roles could also determine how people respond to active shooter incidents.
Therefore, a holistic assessment of countermeasures involving human behavior, people’s daily
roles, and emergency contexts is indispensable to improve human safety during active shooter
incidents. Research questions 3.1 – 3.3 were addressed in this chapter.

102

Chapter 7. Data-driven Crowd Evacuation Simulations for Informing Building
Safety Design
Chapter 7 aims to leverage the empirical findings about human-building-emergency
interactions to develop crowd evacuation simulations for informing building safety design. The
influence of social and environmental factors on people’s wayfinding behavior in building fires
was extracted from Chapter 4 and was modeled using different machine learning and discrete
choice models. Crowd evacuation simulations were further developed to evaluate different
building design options (e.g., visual access levels, number and location of exits and staircases)
with quantitative metrics (i.e., exit choice and evacuation time). The rest of this chapter is
organized as follows. Section 7.1 presents the methodology, including data collection and
extraction, methods to model evacuation decisions, and setups of the evacuation simulation.
Section 7.2 details the model performance in predicting evacuation decisions and simulation
outcomes. Finally, section 7.3 illustrates implications of the findings and section 7.4 concludes
this chapter.
7.1 Methodology
The proposed framework, for the development of the behavioral data-driven agent-based
evacuation simulation, consists of three phases, as shown in Fig. 26. In the first phase, empirical
data is collected via VR-based experiments and factors that affect occupants’ decisions during
emergency evacuation are extracted. In the second phase, different machine learning and discrete
choice models are trained based on the data from the first phase to predict occupants’ wayfinding
decisions. Finally, in the third phase, the trained models are employed in evacuation simulations
with the configuration of agents and buildings. The simulation results are then used for the
evaluation of different building design options.
103

Fig. 26. Proposed framework for the development of behavioral data-driven agent-based
evacuation simulation.
7.1.1 Data collection and extraction
The empirical data used in this study was based on our human-subjects experiment
conducted in 2018 [208,228]. In the experiment, participants were asked to evacuate a virtual
metro station during a fire emergency. The metro station consisted of two floors and three exits.
Initially, participants were placed on the metro platform and a fire broke out in a train approaching
the station shortly after the start of the experiment. Participants then started to evacuate to one of
the exits. The influence of crowd flow on participants’ wayfinding behavior was studied by
including non-player characters (NPCs) who had different directional choices during the
evacuation. Similarly, to introduce the influence of environmental factors, architectural visual
access in the virtual metro station was manipulated at different levels.
To extract the influencing factors and participants’ directional choices during the
evacuation, experiment recordings from the participants’ first-person view were reviewed. The
number of NPCs to each direction when participants made directional choices was manually
counted. Participants’ head rotation and moving direction were used to determine the moment of
their directional choices. Three levels of visual access (i.e., low-1, medium-2, high-3) were defined
104

according to the design manipulations used. Specifically, as shown in Fig. 27, “low” denotes that
no design manipulation was used to increase visual access of a direction; “medium” denotes that
solid walls were replaced with glass walls to increase visual access of a direction; “high” denotes
that in addition to changing wall materials, other measures, including removing columns, moving
the location of ticket booths were used to increase visual access of a direction. Additionally, as the
evacuation involved vertical movement between two floors, a binary variable indicating whether
a direction consisted of vertical movement (staircases) was included in the following data analysis
and evacuation simulation. Demographic attributes (e.g., participants’ age, gender, cultural
backgrounds) were excluded in this study, as these factors were not found to have significant
influence on participants’ directional choices.

Fig. 27. Different levels of visual access in the virtual metro station (a) low, (b) medium, (c) low,
(d) high.
A total of 317 participants from different locations (London, Beijing, and Los Angeles)
were involved in the experiment, among whom the data from 275 participants was used in this
study. Other participants were excluded due to incomplete video recordings of their experiment
sessions. As each participant needed to make two or three consecutive directional choices during
the evacuation, a total of 681 directional choices were extracted from the experiment.
7.1.2 Modeling of evacuation decisions
Directional choices during evacuation can be considered as a classification problem:
predicting whether a direction will be chosen. For each directional choice, direction features were
extracted according to the procedure described in section 7.1.1. The class for each direction was
105

either 0 or 1, denoting whether the direction was chosen by the participant.
Multiple machine learning models were used as the classifier, including k-nearest
neighbors (KNN), support vector machine (SVM), random forest (RF), and neural network (NN).
The KNN model assigns the class of a data point using the most common class among its k-nearest
neighbors. As a hyperparameter, different values of 𝑘 (1, 3, 5, 10) were tested, and 5 was assigned
to 𝑘 . The SVM model learns an optimal hyperplane in the feature space to separate data points
into different classes. With kernel functions, the SVM model can transform features into higher
dimensions and improve the accuracy. Four commonly used kernel functions (linear, quadratic,
cubic, and Gaussian) were used in this study. The RF model builds multiple decision trees and
uses a random subset of features for each tree. The classification output is the class selected by
most trees. Different numbers of trees (10, 50, 100, 200) were tested and eventually 100 decision
trees were used in the RF model. The basic NN model consists of three layers, namely input,
hidden, and output, and neurons in successive layers are connected with edges of different weights.
Neurons in the hidden and output layers can be activated with different activation functions to
allow nonlinear transformation of the input variables. Different numbers of hidden layers and
neurons (1 hidden layer of 50 neurons, 1 hidden layer of 100 neurons, 2 hidden layers of 50 neurons
respectively) were tested and eventually a neural network with one hidden layer of 100 neurons
was used in this study.
The dataset was randomly partitioned into a training set and a testing set. The training set
consisted of 80% data points and was used to train the models. The testing set consisted of the
remaining 20% data points and was used to evaluate the models’ performance. It is important to
note that since our dataset consisted of panel data (i.e., individuals making multiple choices), two
partition methods could be utilized. One is to treat all data points equally and partition them
106

randomly, and the other one is to group data points by individuals before the partition. In this study,
the first method was adopted as it has been commonly used by prior studies [233–235].
Logit models are commonly used for discrete choice modeling and have been applied in a
variety of domains [236,237]. Logit models are based on the assumption that when faced with a
discrete set of options, people are more likely to choose the one of maximal utility [238]. Each
option provides a certain amount of utility as a function of characteristics of the option and the
person making the choice. In this study, a mixed logit model that accommodates heterogeneity in
people’s preferences was employed. The utility provided by option 𝑖 to person 𝑛 can be formulated
as:
𝑈 𝑛𝑖
= 𝛽 𝑛 𝑇 𝑋 𝑛𝑖
+ 𝜀 𝑛𝑖
(1)
where 𝑋 𝑛𝑖
denotes the features of option 𝑖 perceived by person 𝑛 . 𝛽 𝑛 denotes the
preference of person 𝑛 over the features. 𝜀 𝑛𝑖
is an independent and identically distributed error
term. Conditional on 𝛽 𝑛 , the probability of choosing option 𝑖 for person 𝑛 is the standard logit
formula:
𝐿 𝑛𝑖
(𝛽 𝑛 ) =
𝑒 𝛽 𝑛 𝑇 𝑋 𝑛𝑖
∑𝑒 𝛽 𝑛 𝑇 𝑋 𝑛𝑘
𝑘 (2)
Since 𝛽 𝑛 is unknown in advance, the corresponding unconditional probability of choosing
option 𝑖 for person 𝑛 is an integral of 𝐿 𝑛𝑖
(𝛽 𝑛 ) over all possible values of 𝛽 𝑛 :
𝑃 𝑛𝑖
= ∫ (
𝑒 𝛽 𝑇 𝑋 𝑛𝑖
∑𝑒 𝛽 𝑇 𝑋 𝑛𝑘
𝑘 ) 𝑓 (𝛽 |𝜃 ) 𝑑𝛽 (3)
where 𝜃 refers to the collective parameters (e.g., mean and variance) of the distribution of
𝛽 . Accordingly, the log likelihood function can be formulated as:
𝐿𝐿 = ∑ ∑ 𝑦 𝑛𝑖
ln𝑃 𝑛𝑖
𝑖 𝑛 (4)
107

where 𝑦 𝑛𝑖
= 1 if person 𝑛 chooses option 𝑖 and 0 otherwise. As it is computationally
expensive to apply maximum likelihood estimation due to the numerical integration [237],
simulated maximum likelihood method with 1,000 Halton draws was used to estimate the model
parameters that maximize the simulated log likelihood.
Additionally, to examine how different influencing factors and emergency contexts affect
occupants’ decision-making processes and thereby the simulation results, the mixed logit model
from Haghani and Sarvi [55] (thereafter referred to as HS model) was used in the evacuation
simulation and compared with the trained machine learning and discrete choice models in this
study. The HS model was estimated based on evacuation experiments in a laboratory environment,
where 150 participants performed evacuation trials in a temporary setup with different number and
spatial distribution of exits. The experiment recordings were analyzed by the authors to extract
participants’ decisions along with the influencing factors. The variables included in the HS model
were distance to each exit, exit visibility, number of crowds near each exit, and number of crowds
moving to visible/invisible exits. Exit visibility in the HS model was defined as whether the exit
itself is directly visible from the person’s location, which is different from visual access defined
in this study (i.e., visual access resulted from architectural attributes). Reasons for choosing the
HS model for the comparison are that (1) the HS model is a type of discrete choice model and
considered both social and environmental factors and (2) the HS model was based on empirical
data from a distinct scenario. Contextual factors vary across different building emergencies
scenarios and their impact on employing evacuation simulations for building safety design remains
underexplored. Hence, the comparison could help to identify the critical factors for the
applicability of evacuation simulations.
7.1.3 Evacuation simulation
108

Agents in the simulation were autonomous in perceiving the environment and making
decisions to evacuate. At the strategic level, consistent with our human-subjects experiment, it was
assumed that there were sufficient environmental cues for agents to perceive the emergency and
start to evacuate. Therefore, each agent was randomly assigned a short pre-evacuation time
uniformly ranging from 0 to 10 seconds.
At the tactical level, three models, namely the neural network model, the mixed logit model
trained using our experiment data, and the HS model were employed by agents. The reason for
choosing the neural network model was that its prediction accuracy was higher than the other
machine learning models. The agents had a view angle of 120° and they could perceive the number
of crowds moving to each direction in their field of view [239]. Other features of each direction
could also be perceived and considered by the agents when making directional choices. Visual
access level and whether a direction consists of vertical movement were annotated in the building
design. Distance to exits, exit visibility, and number of crowds near exits, which were influencing
factors included in the HS model, were calculated in real time during the simulation. One thing to
be pointed out is that while there were multiple decision points in the evacuation simulation,
distance to exits was measured between the agents’ location and the final exits, which was
consistent with the setup in [55].
At the operational level, social force model was used for agents’ stepwise movement.
Specifically, once agents made tactical-level decisions, they would calculate the shortest weighted
path to the destination considering the presence of other agents and obstacles. The next point on
the calculated path was then fed into the social force model as the desired force. Forces imposed
by other agents and walls were also calculated in real time and the combined effect was used to
update the agents’ speed magnitude and direction. The parameters in the social force model was
109

set in accordance with the ones in prior studies [46,240].
A two-floor building (a metro station) with three different layouts was used for the
evacuation simulation, as shown in Fig. 28. Layout 1 is analogous to the metro station layout used
in our human-subjects experiment. The difference between layouts 1 & 2 is that in layout 2, Exit
1 is closer to and visible from decision point (DP) 2. For layout 3, there are 4 exits in contrast to 3
exits in layouts 1 & 2, and agents would choose among 3 exits at DP 3. For each building layout,
multiple directional choices need to be made to navigate from the starting area to an exit.
Specifically, from the starting area/DP 1, agents need to make the first directional choice between
the hallway on the left and Staircase 1 on the right. If agents choose the hallway and arrive at DP
2, they need to make another directional choice between going to Staircase 2 on the left and keep
moving in the hallway to Exit 1. On the other hand, if agents choose Staircase 1 or Staircase 2,
they will be on the underground floor and further navigate to DP 3. At this point, agents will make
another directional choice to go to the exits (Exits 2 & 3 in layouts 1 & 2, Exits 2-4 in layout 3).
Moreover, to examine the influence of visual access on the simulation results, for each layout,
different visual access levels were assigned to the direction options, as presented in Table 10.

110

Fig. 28. Building layouts used in the evacuation simulation (walls are shown in white lines,
staircases are shown in gray rectangles, and arrows denote the moving direction on staircases.
Agents can only access the areas bounded by outer walls).
Table 10. Summary of building design scenarios used in the evacuation simulation.
At the beginning of the simulation, 100 agents were concurrently generated within the
starting area with random location and head orientation. The agents make directional choices
according to the mechanism explained above. One simulation run ends when all agents have
reached an exit. Both evacuation time and exit choice were recorded and used as evacuation
metrics. For each scenario, the simulation was repeated for 50 times.
7.2 Results
7.2.1 Model performance
Four performance metrics, namely accuracy, precision, recall, and F1 score were used for
evaluating the models’ performance. One can see from Table 11 that overall, all models had
adequate performance. Considering all the metrics collectively, the neural network and mixed logit
models had slightly higher performance than other models. Therefore, these two models were
further used for the evacuation simulation.

No. Layout No. of exits No. of staircases
Visual access
DP2 Staircase 1 Staircase 2 Exit1 Exit2 Exit3 Exit4
1 1 3 4 1 1 1 1 1 1 N/A
2 1 3 4 1 1 1 2 2 1 N/A
3 1 3 4 1 2 3 1 3 1 N/A
4 1 3 4 2 1 2 1 1 2 N/A
5 2 3 4 1 1 1 2 1 1 N/A
6 2 3 4 1 1 1 3 2 1 N/A
7 2 3 4 1 2 3 2 3 1 N/A
8 2 3 4 2 1 2 2 1 2 N/A
9 3 4 5 1 1 1 2 1 1 1
10 3 4 5 1 1 2 2 2 1 1
11 3 4 5 1 2 3 2 1 2 1
12 3 4 5 2 1 2 3 3 2 1
111

Table 11. Performance of the examined models.
Metrics
Model
KNN Linear SVM Quadratic SVM Cubic SVM Gaussian SVM RF NN MLM
Accuracy 79.56% 83.94% 83.94% 83.21% 83.21% 79.56% 87.59% 86.86%
Precision 84.62% 90.77% 96.92% 98.46% 86.15% 87.69% 92.31% 87.69%
Recall 75.34% 78.66% 75.90% 74.42% 80.00% 74.03% 83.33% 85.07%
F1 score 79.71% 84.29% 85.13% 84.77% 82.96% 80.28% 87.59% 86.36%
Note: KNN-k nearest neighbors, SVM-support vector machine, RF-random forest, NN-neural
network, MLM-mixed logit model.
After determining the models to be used in the evacuation simulation, the neural network
and mixed logit models were retrained using the entire dataset. Estimation of the mixed logit
model’s parameters is presented in Table 12. Goodness-of-fit measures how well the observed data
matches the expected data under the model. McFadden’s pseudo R
2
is a commonly used goodness-
of-fit metric, and a value above 0.2 usually suggests satisfactory model fit [241]. As shown in
Table 12, the value of McFadden’s pseudo R
2
is 0.415, which suggests a sufficiently good model
fit. Apart from goodness-of-fit, significant p-values for the estimated mean of model parameters
revealed that all the three factors had significant influence on directional choices. The negative
sign of whether a direction consists of vertical movement indicated that participants tended to
avoid vertical movement during the evacuation. Likewise, the positive sign of visual access of a
direction and number of crowds moving to a direction indicated that higher visual access and
crowd flow increased the attractiveness of a direction. This result is consistent with the findings in
our experiment [208,228] but contradictory to the findings in [55] to certain extent as no significant
following behavior was found in [55]. The possible reason is that in [55], the building complexity
was lower and participants were familiar with the layout as they performed evacuation trials
multiple times. As a result, the participants were less inclined to follow the crowd to avoid
congestions. Estimation of the standard deviations showed that despite the overall preference, there
existed heterogeneities in individual’s attitudes towards vertical movement and crowd flow, as
reflected by the significant p-values. On the contrary, individual’s preference over visual access
112

was rather consistent (p-value > 0.05).
Table 12. Estimation of the mixed logit model’s parameters.
Model parameters β Standard error p-value
Mean
Whether the direction consists of vertical
movement
-1.675 0.522 0.001 ***
Visual access of the direction 0.626 0.279 0.025 **
Number of crowds moving to the direction 0.536 0.150 < 0.001 ***

Standard deviation
Whether the direction consists of vertical
movement
3.085 1.071 0.004 ***
Visual access of the direction 0.343 2.308 0.882
Number of crowds moving to the direction 0.247 0.104 0.018 ***

Goodness of fit
Log likelihood at constant -472.03
Log likelihood at convergence -276.29
McFadden’s pseudo R
2
0.415
7.2.2 Simulation outcome
First, the exit choice was analyzed as it reflects how agents employed the models to make
directional choices, which fundamentally determined agents’ evacuation path and performance.
Fig. 29 shows exit choices of agents using the neural network and mixed logit models in each
building design scenario as specified in Table 10. For agents using the HS model, since the
influence of visual access was not included, the simulation was only run for each building layout
without enumerating all the building design scenarios. Exit choices of agents using the HS model
are shown in Fig. 30.
To examine whether building designs elicited different agent behavior, Chi-square test was
conducted for each layout across different scenarios. For both the neural network and mixed logit
models, there were (marginally) significant differences across building design scenarios (𝜒 2
>=
16.774, P-value <= 0.052). Based on the estimated parameters of the mixed logit model and
findings in [208,228], agents using both models made reasonable exit choices and generated
similar results. For instance, in scenarios 1 & 2, Exit 1 was chosen by most agents as Exit 1 was
113

on the same floor as the starting area and the visual access of Exit 1 was improved in scenario 2.
By contrast, in scenario 3, visual access of Staircase 2 was increased while the visual access of
Exit 1 was decreased, hence the number of agents choosing Exit 1 significantly reduced. Moreover,
in scenario 3, Exit 2 had much higher visual access than Exit 3 (visual access 3 vs. 1), which was
also reflected by the unbalanced choices between Exits 2 & 3. For scenarios 5-8 that corresponded
to layout 2 and scenarios 9-12 that corresponded to layout 3, the two models also captured the
differences in building design and produced comparable results.

114

Fig. 29. Exit choices of agents using the neural network and mixed logit models for each
building design scenario.
To examine whether there were any differences between the neural network and mixed
logit models, Chi-square test was conducted. As shown in Table 13, there existed several
discrepancies in the two models’ simulation results. First, the mixed logit model could capture
more subtle building design differences. For instance, visual access of Exit 2 improved from 1 to
2 from scenario 1 to scenario 2, and visual access of Exit 3 remained as 1. While not statistically
different, agents using the mixed logit model chose Exit 2 more frequently in scenario 2, whereas
agents using the neural network model had almost identical choices in scenarios 1 & 2. Second,
when the available directions provided mixed information, such as in scenarios 3, 7 & 8 and 10 &
11 where staircases had high visual access and hallways had low visual access, agents using the
neural network and mixed logit models exhibited more significantly different behaviors. Third, the
results generated by the neural network model may lack interpretability. For instance, it is expected
that Exit 2 was more frequently chosen than Exit 3 due to higher visual access in scenario 7.
However, it is unclear why agents using the neural network model preferred Exit 2 over Exit 1, as
Exit 1 also had medium level of visual access and was on the same floor as the starting area. These
results revealed that while the neural network model outperformed the mixed logit model by a
small margin with regards to prediction accuracy, it may fall short in explanatory power, as
suggested by prior studies [233,235].
Table 13. Comparison of the neural network and mixed logit models in agents’ exit choices.
Scenario 𝜒 2
p-value
1 0.611 0.737
2 3.157 0.206
3 4.754 0.093 *
4 1.924 0.382
5 2.481 0.289
6 4.086 0.130
7 14.182 0.001 ***
8 7.967 0.019 **
115

9 1.583 0.663
10 12.989 0.005 **
11 9.812 0.020 **
12 0.840 0.840
Exit choices by agents using the HS model are shown in Fig. 30. The agents’ exit choices
were significantly different (all p-value < 0.05) from the agents using the neural network and
mixed logit models. In scenarios 1-4, Exit 2 was chosen by most agents followed by Exit 3,
whereas Exit 1 was rarely chosen. In scenarios 5-8, the number of agents choosing Exits 2 & 3
was almost the same and the number of agents choosing Exit 1 increased. In scenarios 9-12, on
the other hand, Exit 1 became the most chosen exit, followed by Exits 2 & 3, and then Exit 4.
Possible reasons for such distinct exit choices are as follows. In the HS model, agents would prefer
closer exits. Since the HS model was estimated based on a laboratory setup, the building scale is
much smaller than the virtual metro station used in our experiment. As a result, exit locations
largely affected the distance between agents and exits and influenced agents’ exit choices. In
scenarios 1-4, Exit 1 is far away from DP 1, hence most agents chose Staircase 1 when making the
first directional choice and evacuated via Exits 2 & 3. This also explains the unbalanced usage of
Exits 2 & 3: Most agents chose Staircase 2 at DP 1, hence when they navigated to DP 3 on the first
floor, they were closer to Exit 2 and were more likely to choose it. In scenarios 5-8, the distance
between Exit 1 and the starting area is much shorter, which motivated more agents to evacuate via
Exit 1. In scenarios 9-12, as Exit 4 is far away from the starting area, even more agents chose to
evacuate via Exit 1. Moreover, very few agents chose Exit 4 since Exit 4 is further away from DP
3 compared with Exits 2 & 3.
116

Fig. 30. Exit choices of agents using the HS model for each building layout.
For agents using the neural network and mixed logit models, the approximated probability
density of their evacuation time is shown in Fig. 31, which was plotted using the Gaussian kernel
density estimation. As can be seen, there are two local maxima in the density function for scenarios
5-12, which is related to the difference between layout 1 and layouts 2 & 3: The location of Exit 1
is much closer to the starting location than Exits 2 & 3 in layouts 2 & 3, hence choosing Exit 1 vs.
Exits 2 & 3 would lead to apparent difference in evacuation times. Similar to exit choices, the two
models produced comparable results in general, even though discrepancies existed in several
scenarios. For instance, in scenarios 7, 10, and 11, the evacuation time for agents using the mixed
logit model had much higher probability density around 40-50 seconds compared with agents using
the neural network model. This finding is aligned with agents’ exit choices. As illustrated in Fig.
29, in scenarios 7, 10, and 11, more agents using the mixed logit model chose Exit 1 than agents
using the neural network model.
117

Fig. 31. Evacuation time of agents using the neural network and mixed logit models for each
building design scenario.
Similarly, Fig. 32 shows the approximated probability density of evacuation time for agents
using the HS model. In scenarios 1-4, agents predominantly chose Exits 2 & 3. As it takes longer
to evacuate via Exits 2 & 3 than Exit 1, the evacuation time had high probability density
approximately between 60 seconds and 90 seconds. In scenarios 5-8 and scenarios 9-12, since
there were more agents choosing Exit 1, they could evacuate the building in shorter times, which
118

resulted in the peak around 40-50 seconds in the approximated probability density.

Fig. 32. Evacuation time of agents using the HS model for each building layout.
7.3 Discussion
The modeling and simulation results revealed that visual access, crowd flow, and vertical
movement all had significant impacts on directional choices and overall evacuation performance.
At the individual level, parameters of the mixed logit model revealed that directions with higher
visual access and crowd flow could provide higher utilities to occupants, whereas vertical
movement decreased the attractiveness of a direction. At the crowd level, the simulation results
showed that agents considered all factors collectively when making directional choices. According
to the simulation results in the 12 scenarios, while the agents tended to avoid vertical movement
and chose the direction on the same floor, increasing visual access of staircases motivated the
agents to move between floors. Since crowd flow had a positive impact on directional choice,
agents who started their evacuation late tended to follow others in front and choose staircases as
well. This finding suggests that instead of tuning parameters for a single attribute (e.g., number or
location of exits), the interaction effect of multiple building attributes should be evaluated
collectively in building safety design, as echoed by several prior studies [27,46]. For instance,
when manipulating visual access, building layout and exit location should be taken into
consideration.
119

By comparing the machine learning and mixed logit models, it was found that all the
models achieved adequate prediction accuracy. The neural network model achieved the best
performance followed by the mixed logit model. This finding is accordant with several prior
studies that found machine learning models were more advantageous in prediction accuracy than
logit models [233,235,242]. Under more complex scenarios, the neural network model may
outperform the mixed logit model by a bigger margin due to its capacity of capturing inherent
nonlinearities in the training data. The two models were further evaluated based on the simulation
results. It was found that while the two models were able to respond to different design options
reasonably and generated comparable results, there still existed discrepancies in their simulation
outcomes. Output of the mixed logit model was well-aligned with its parameter estimation.
However, for the neural network model, some output may lack interpretability. As suggested by
prior studies, logit models typically focus on fitting the model to training data and guaranteeing
the model’s interpretability [233,243]. On the other hand, machine learning models prioritize
prediction accuracy and are rarely used to extract behavioral insights. Even though several metrics,
such as variable importance, direction of influence, and partial dependence plots, have been used
in prior studies to interpret machine learning models, it was still found that logit models had better
explanatory power [233,244]. Therefore, there should be tradeoffs between models’ prediction
accuracy and interpretability, and different models could be used to complement each other. For
instance, machine learning models can be used to identify better specifications for logit models
due to the capability of capturing nonlinearities in data automatically [233].
In addition, the comparison of the trained models in this study and the HS model showed
that agents using the HS model had significantly different behavior from agents using the neural
network and mixed logit models. The reason was that the HS model was estimated based on
120

evacuation experiments in different setups with regards to building complexity, level of
crowdedness, participants’ familiarity with the building, and so on. First, since the building layout
was relatively simple and participants were familiar with the layout, they could rely less on
surrounding people and hence there was no significant following tendency [227]. Second, the size
of building used in [55] was much smaller than the building used in our evacuation simulation.
Therefore, when manipulating exit locations, the difference of distance between exits and agents
could cause significant perturbations to the HS model. Similarly, when applying the trained models
in this study to drastically distinct scenarios, different outputs could be expected. The usage of
data-driven evacuation simulations should be closely linked to both the development and
application scenarios. Factors related to occupants, buildings, and emergencies could all affect
simulation results and hence the applicability of evacuation simulations. Even though evacuation
simulations are not meant to be used in identical building emergency scenarios as where they are
developed, the findings in this study revealed that it is important to choose an evacuation
simulation that corresponds to the target scenario at certain levels (e.g., building size, level of
crowdedness, etc.).
Based on the findings in this study, four recommendations are proposed for developing
evacuation simulations for building safety design. First, future studies may carefully consider what
influencing factors to include in evacuation simulations. While most prior evacuation simulations
focused on obviously relevant building attributes (e.g., size and number of exits, corridor width),
recent studies showed that other building attributes (e.g., wall transparency, landmarks) can also
influence occupant behavior during building emergencies [245], which indicates the necessity to
include these factors in evacuation simulation models. That being said, researchers also found that
introducing more variables may increase the difficulty of model calibration and applying the model
121

to different scenarios [156]. Therefore, a balance between model complexity and
comprehensiveness should also be considered. Second, as mentioned above, most prior studies
leveraged evacuation simulations for building safety design and concluded their findings without
considering the dependency on specific building and emergency scenarios. As the performance of
building safety design is contingent on different types of building-, emergency-, and occupant-
characteristics, more cross validations are suggested for building safety design. Even though it is
a well-recognized challenge to develop robust evacuation simulations that ensures generalizability
to a wide variety of scenarios [44], researchers are suggested to present the findings of evacuation
simulation results in relation to clearly defined building emergency scenarios. Third, as multiple
metrics may be collectively considered for building design evaluation, it is suggested to adopt an
evacuation simulation with high explanatory power to associate the metrics with the examined
scenarios to inform the improvement of building safety design. Fourth, different models for
representing occupant behavior should be further investigated. For instance, recent studies coupled
reinforcement learning with social force model and logit models [246,247]. Such alternative
models may be especially helpful in building emergency scenarios (e.g., active shooter incidents)
that involve people in different roles with different goals and behavioral patterns [219,248]. Finally,
this study also has its limitations: The findings were based on the data collected in our human-
subjects experiments, which could not possibly cover the whole spectrum of building emergency
scenarios. For instance, while it did not occur in our experiments, some occupants may not be able
to evacuate the building in real-world emergencies. Effective approaches to represent these
occupants in evacuation simulations and consider them in building safety design should be further
explored. Moreover, while various social and environmental factors were found to collectively
affect occupant behavior and the outcome of building emergencies, only three including factors
122

were examined in this study. More factors that affect occupant behavior and their interaction
effects could be explored in future research.
7.4 Conclusion
In this study, data-driven agent-based evacuation simulations were developed for the
evaluation of building design options. The influences of crowd flow, visual access, and vertical
movement on human behavior during building emergencies were extracted from human-subjects
experiments and examined. Both machine learning and mixed logit models were trained to predict
people’s directional choices during emergency evacuation. All the models achieved adequate
prediction accuracy, and the results were congruous with prior studies. The neural network and
mixed logit models trained in this study, as well as a mixed logit model from a prior study, were
incorporated in crowd evacuation simulations to evaluate different building design options. The
simulation results revealed that different building attributes could collectively affect agents’
directional choices and evacuation performance. While both the trained neural network and mixed
logit models generated comparable results, the mixed logit model was more advantageous in its
interpretability. By comparing the trained models in this study with another model in a prior study,
significant differences in simulation results were found since the models were based on different
building emergency scenarios and considered different influencing factors. The results revealed
that building size, level of crowdedness, and participants’ familiarity with the building could all
affect the applicability of crowd evacuation simulations for building safety design. Future studies
are recommended to consider what influencing factors to include in evacuation simulations,
examine the robustness and applicability of simulation results more closely, and explore different
modeling methods to represent human behavior during building emergencies. Research questions
4.1 – 4.2 were addressed in this chapter.
123

Chapter 8. Conclusions and Future Directions
This dissertation contributes to deepening our understanding of human-building-
emergency interactions via the use of various research methods (human-subject experiments, focus
group interviews, mathematical modeling, and crowd evacuation simulations) and leveraging the
knowledge for informing building safety design and emergency management. The work in Chapter
4 helps to understand how visual access and people’s cultural background influence their
wayfinding behavior during building fires by conducting virtual reality-based human-subjects
experiments in multiple locations. The results revealed that increasing visual access could improve
people’s evacuation performance and reduce their tendency of following others. Since active
shooter incidents present an increasing threat to the American society and human behavior during
building emergencies is highly correlated to emergency scenarios, Chapter 5 and Chapter 6 aim to
identify and evaluate a variety of security countermeasures that have been proposed and used in
buildings in response to the risk of active shooter incidents. In Chapter 5, a series of focus group
interviews were conducted with domain experts to identify security countermeasures for active
shooter incidents and critical considerations for the implementation of countermeasures. The
results in Chapter 5 revealed that certain countermeasures (e.g., access control) may be beneficial
for building security but have negative impacts on human behavior when an active shooter incident
occurs. Therefore, in Chapter 6, human-subject experiments were conducted to empirically assess
the influence of countermeasures on people’s responses to active shooter incidents in office and
school buildings. Other influencing factors, including people’s occupational background, building
and social contexts, were also investigated in Chapter 6. The results in Chapter 6 suggested that
with the implementation of countermeasures, people’s response time and decisions during active
shooter incidents would be significantly affected. Occupational background and building type
124

would also affect people’s responses and perceptions. To leverage the obtained knowledge about
human-building-emergency interactions for informing building safety design and emergency
management, in Chapter 7, we modeled the influence of social (i.e., crowd flow) and
environmental (i.e., visual access and vertical movement) factors on people’s wayfinding behavior
during building fires based on the experiment results in Chapter 4. Machine learning and discrete
choice models were employed to predict people’s evacuation decisions and were further integrated
in crowd evacuation simulations for evaluating different building design options (e.g., number and
location of exits, staircases, and visual access). The results in Chapter 7 revealed that both machine
learning and discrete choice models could accurately predict people’s evacuation decisions,
whereas discrete choice models have better interpretability. Different building attributes could
collectively impact exit choices and evacuation time. Building size, crowd density, and people’s
familiarity with the building are critical factors for the applicability of crowd evacuation
simulations.
There are several limitations in this dissertation that require future investigations. First,
while we created virtual environments to represent emergency scenarios such as building fires and
active shooter incidents, the virtual environments cannot affect participants’ physical mobility and
no thermal and olfactory stimuli were provided, which might not evoke participants’ response at
the same level as real-world emergencies. To enhance the sense of presence that participants
experience in virtual environments, future studies could provide more stimuli channels (e.g.,
thermal, olfactory, and haptic feedback) to make the virtual emergency scenario more comparable
to real-world emergencies. Second, the results in this dissertation are based on data-driven
approaches (both human-subject experiments and focus group interviews), hence the findings are
dependent on characteristics of the participants. For instance, the participants in human-subject
125

experiments (Chapter 4 and Chapter 6) were from certain cultural and occupational backgrounds.
Future studies are suggested to conduct investigations on more diverse population groups (e.g.,
elderly, people with mobility impairment). Third, for the development of crowd evacuation
simulations presented in Chapter 7, one limitation lies in the fact that they were based on the
experiments in Chapter 4, hence the application is constrained by the investigated phenomenon
and variables collected in the experiments. Future research could focus on emergency scenarios
that involve more complex variables (e.g., people with different roles). That being said, the
framework of developing the crowd evacuation simulations, including data collection, exploratory
analysis of data, model selection and training, and simulation configurations could provide useful
insights for future research.

126

Publications
Peer-reviewed journal papers (published)
R. Zhu, J. Lin, B. Becerik-Gerber, N. Li, Human-building-emergency interactions and their
impact on emergency response performance: A review of the state of the art, Saf. Sci. 127 (2020)
104691. doi:10.1016/j.ssci.2020.104691 (Chapter 2)

R. Zhu, J. Lin, B. Becerik-Gerber, N. Li, Influence of architectural visual access on emergency
wayfinding: A cross-cultural study in China, United Kingdom and United States, Fire Saf. J. 113
(2020) 102963. doi:10.1016/j.firesaf.2020.102963 (Chapter 4)

R. Zhu, G.M. Lucas, B. Becerik-Gerber, E.G. Southers, Building preparedness in response to
active shooter incidents: Results of focus group interviews, Int. J. Disaster Risk Reduct. 48 (2020)
101617. doi:10.1016/j.ijdrr.2020.101617 (Chapter 5)

R. Zhu, G.M. Lucas, B. Becerik-Gerber, E.G. Southers, E. Landicho, The impact of security
countermeasures on human behavior during active shooter incidents, Sci. Rep. 12 (2022) 1–15.
doi:10.1038/s41598-022-04922-8 (Chapter 6)

J. Lin, R. Zhu, N. Li, B. Becerik-gerber, How occupants respond to building emergencies: A
systematic review of behavioral characteristics and behavioral theories, Saf. Sci. 122 (2020)
104540. doi:10.1016/j.ssci.2019.104540

J. Lin, R. Zhu, N. Li, B. Becerik-Gerber, Do people follow the crowd in building emergency
evacuation? A cross-cultural immersive virtual reality-based study, Adv. Eng. Informatics. 43
(2020) 101040. doi:10.1016/j.aei.2020.101040

M. Awada, R. Zhu, B. Becerik-Gerber, G.M. Lucas, E.G. Southers, An integrated emotional and
physiological assessment for VR-based active shooter incident experiments, Adv. Eng.
Informatics. 47 (2021) 101227. doi:10.1016/j.aei.2020.101227

Peer-reviewed journal papers (under review)
R. Zhu, B. Becerik-Gerber, J. Lin, N. Li, Behavioral data-driven agent-based evacuation
simulation for building safety design using machine learning and discrete choice models, Adv.
Eng. Informatics (Chapter 7)

127

R. Liu, R. Zhu, B. Becerik-Gerber, G.M. Lucas, E.G. Southers, Be prepared: How training and
emergency type affect evacuation behavior, J. Comput. Assist. Learn.
Peer-reviewed conference papers (published)
R. Zhu, B. Becerik-Gerber, J. Lin, N. Li, Modeling the impact of visual access and crowd flow on
emergency wayfinding: From empirical investigations to simulations, ASCE International
Conference on Computing in Civil Engineering (2021), Orlando, FL.

R. Zhu, B. Becerik-Gerber, G.M. Lucas, E.G. Southers, D.V. Pynadath, Information and
representation requirements for virtual environments to study human-building interactions in
active shooter incidents. ASCE International Conference on Computing in Civil Engineering
(2019), Atlanta, GA.

R. Zhu, J. Lin, B. Becerik-Gerber, N. Li, Virtual reality-based studies of human emergency
behavior in built environments: A systematic review. International Conference on Construction
Applications of Virtual Reality (2018), Auckland, New Zealand.

128

References
[1] The U.S. Fire Administration, U.S. Fire Statistics, (2020).
https://www.usfa.fema.gov/data/statistics/.
[2] Federal Bureau of Investigation, Active Shooter Incidents in the United States from 2000-
2018, (2018). https://www.fbi.gov/about/partnerships/office-of-partner-
engagement/active-shooter-incidents-graphics.
[3] N.E. Klepeis, W.C. Nelson, W.R. Ott, J.P. Robinson, A.M. Tsang, P. Switzer, J. V. Behar,
S.C. Hern, W.H. Engelmann, The National Human Activity Pattern Survey (NHAPS): A
resource for assessing exposure to environmental pollutants, J. Expo. Anal. Environ.
Epidemiol. 11 (2001) 231–252. doi:10.1038/sj.jea.7500165.
[4] R. Zhu, J. Lin, B. Becerik-Gerber, N. Li, Human-building-emergency interactions and
their impact on emergency response performance: A review of the state of the art, Saf. Sci.
127 (2020) 104691. doi:10.1016/j.ssci.2020.104691.
[5] J. Lin, R. Zhu, N. Li, B. Becerik-gerber, How occupants respond to building emergencies:
A systematic review of behavioral characteristics and behavioral theories, Saf. Sci. 122
(2020) 104540. doi:10.1016/j.ssci.2019.104540.
[6] M. Haghani, M. Sarvi, Heterogeneity of decision strategy in collective escape of human
crowds: On identifying the optimum composition, Int. J. Disaster Risk Reduct. 35 (2019)
101064. doi:10.1016/j.ijdrr.2019.101064.
[7] X. Pan, C.S. Han, K. Dauber, K.H. Law, Human and social behavior in computational
modeling and analysis of egress, Autom. Constr. 15 (2006) 448–461.
doi:10.1016/j.autcon.2005.06.006.
[8] M. Kobes, I. Helsloot, B. de Vries, J.G. Post, Building safety and human behaviour in fire:
A literature review, Fire Saf. J. 45 (2010) 1–11. doi:10.1016/j.firesaf.2009.08.005.
[9] R. Passini, Wayfinding: A conceptual framework, Urban Ecol. 5 (1981) 17–31.
doi:10.1016/0304-4009(81)90018-8.
[10] E. Vilar, F. Rebelo, P. Noriega, E. Duarte, C.B. Mayhorn, Effects of competing
environmental variables and signage on route-choices in simulated everyday and
emergency wayfinding situations, Ergonomics. 57 (2014) 511–524.
doi:10.1080/00140139.2014.895054.
[11] C.H. Tang, W.T. Wu, C.Y. Lin, Using virtual reality to determine how emergency signs
facilitate way-finding, Appl. Ergon. 40 (2009) 722–730.
doi:10.1016/j.apergo.2008.06.009.
[12] M.T. Kinateder, B. Comunale, W.H. Warren, Exit choice in an emergency evacuation
scenario is influenced by exit familiarity and neighbor behavior, Saf. Sci. 106 (2018) 170–
175. doi:10.1016/j.ssci.2018.03.015.
129

[13] R. Lovreglio, A. Fonzone, L. dell’Olio, D. Borri, A study of herding behaviour in exit
choice during emergencies based on random utility theory, Saf. Sci. 82 (2016) 421–431.
doi:10.1016/j.ssci.2015.10.015.
[14] M. Isobe, D. Helbing, T. Nagatani, Experiment, theory, and simulation of the evacuation
of a room without visibility, Phys. Rev. E. 69 (2004) 10.
doi:10.1103/PhysRevE.69.066132.
[15] W. Yuan, K.H. Tan, A model for simulation of crowd behaviour in the evacuation from a
smoke-filled compartment, Phys. A Stat. Mech. Its Appl. 390 (2011) 4210–4218.
doi:10.1016/j.physa.2011.07.044.
[16] M. Haghani, M. Sarvi, Z. Shahhoseini, M. Boltes, How simple hypothetical-choice
experiments can be utilized to learn humans’ navigational escape decisions in
emergencies, PLoS One. 11 (2016) 1–24. doi:10.1371/journal.pone.0166908.
[17] G.Y. Jeon, J.Y. Kim, W.H. Hong, G. Augenbroe, Evacuation performance of individuals
in different visibility conditions, Build. Environ. 46 (2011) 1094–1103.
doi:10.1016/j.buildenv.2010.11.010.
[18] R. Lovreglio, V. Gonzalez, R. Amor, M. Spearpoint, J. Thomas, M. Trotter, R. Sacks, The
Need for Enhancing Earthquake Evacuee Safety by Using Virtual Reality Serious Games,
Lean Comput. Constr. Congr. - Vol. 1 Proc. Jt. Conf. Comput. Constr. (2017) 381–389.
doi:10.24928/JC3-2017/0058.
[19] J.P. Blair, K.W. Schweit, A study of active shooter incidents in the United States between
2000 and 2013, Washington, DC, 2014.
[20] M. Haghani, M. Sarvi, Crowd behaviour and motion: Empirical methods, Transp. Res.
Part B Methodol. 0 (2017) 1–42. doi:10.1016/j.trb.2017.06.017.
[21] G. Bernardini, M. D’Orazio, E. Quagliarini, Towards a “behavioural design” approach for
seismic risk reduction strategies of buildings and their environment, Saf. Sci. 86 (2016)
273–294. doi:10.1016/j.ssci.2016.03.010.
[22] G. Jeon, W. Hong, Characteristic features of the behavior and perception of evacuees from
the Daegu subway fire and safety measures in an underground fire, J. Asian Archit. Build.
Eng. 8 (2009) 415–422. doi:10.3130/jaabe.8.415.
[23] B.E. Aguirre, M.R. Torres, K.B. Gill, H. Lawrence Hotchkiss, Normative Collective
Behavior in The Station Building Fire, Soc. Sci. Q. 92 (2011) 100–118.
doi:10.1111/j.1540-6237.2011.00759.x.
[24] A.R. Mawson, Understanding Mass Panic and Other Collective Responses to Threat and
Disaster, Psychiatry Interpers. Biol. Process. 68 (2005) 95–113.
doi:10.1521/psyc.2005.68.2.95.
[25] M. Moussaïd, M. Trauernicht, Patterns of cooperation during collective emergencies in
130

the help-or-escape social dilemma, Sci. Rep. 6 (2016) 33417. doi:10.1038/srep33417.
[26] X. Pan, C.S. Han, K. Dauber, K.H. Law, A multi-agent based framework for the
simulation of human and social behaviors during emergency evacuations, AI Soc. 22
(2007) 113–132. doi:10.1007/s00146-007-0126-1.
[27] V. Ha, G. Lykotrafitis, Agent-based modeling of a multi-room multi-floor building
emergency evacuation, Phys. A Stat. Mech. Its Appl. 391 (2012) 2740–2751.
doi:10.1016/j.physa.2011.12.034.
[28] S. Gwynne, E.R. Galea, P.J. Lawrence, L. Filippidis, Modelling occupant interaction with
fire conditions using the buildingEXODUS evacuation model, Fire Saf. J. 36 (2001) 327–
357.
[29] L. Cao, J. Lin, N. Li, A virtual reality based study of indoor fire evacuation after active or
passive spatial exploration, Comput. Human Behav. 90 (2019) 37–45.
doi:10.1016/j.chb.2018.08.041.
[30] G. Bernardini, E. Quagliarini, M. D’Orazio, Towards creating a combined database for
earthquake pedestrians’ evacuation models, Saf. Sci. 82 (2016) 77–94.
doi:10.1016/j.ssci.2015.09.001.
[31] C.E. Rusbult, P.A.M. Van Lange, Interdependence, Interaction, and Relationships, Annu.
Rev. Psychol. 54 (2002) 351–375. doi:10.1146/annurev.psych.54.101601.145059.
[32] E.L. Quarantelli, The Nature and Conditions of Panic, Am. J. Sociol. 60 (n.d.) 267–275.
doi:10.2307/2772684.
[33] L. Clarke, panic : myth or reality ?, Contexts. 1 (2002) 21–26.
[34] J.D. Sime, Affiliative behaviour during escape to building exits, J. Environ. Psychol. 3
(1983) 21–41. doi:10.1016/S0272-4944(83)80019-X.
[35] J. Drury, C. Cocking, S. Reicher, Everyone for themselves? A comparative study of crowd
solidarity among emergency survivors, Br. J. Soc. Psychol. 48 (2009) 487–506.
doi:10.1348/014466608X357893.
[36] D. Nilsson, A. Johansson, Social influence during the initial phase of a fire evacuation-
Analysis of evacuation experiments in a cinema theatre, Fire Saf. J. 44 (2009) 71–79.
doi:10.1016/j.firesaf.2008.03.008.
[37] S. Heliövaara, J.M. Kuusinen, T. Rinne, T. Korhonen, H. Ehtamo, Pedestrian behavior
and exit selection in evacuation of a corridor - An experimental study, Saf. Sci. 50 (2012)
221–227. doi:10.1016/j.ssci.2011.08.020.
[38] M.T. Kinateder, W.H. Warren, K.B. Schloss, What color are emergency exit signs? Egress
behavior differs from verbal report, Appl. Ergon. 75 (2019) 155–160.
doi:10.1016/j.apergo.2018.08.010.
131

[39] J. Coaffee, C. Moore, D. Fletcher, L. Bosher, Resilient design for community safety and
terror-resistant cities, Proc. Inst. Civ. Eng. - Munic. Eng. 161 (2008) 103–110.
doi:10.1680/muen.2008.161.2.103.
[40] S. Khiun Then, M. Loosemore, Terrorism prevention, preparedness, and response in built
facilities, Facilities. 24 (2006) 157–176. doi:10.1108/02632770610665766.
[41] Federal Bureau of Investigation, Responding to an Active Shooter Crisis Situation, (2016).
https://www.fbi.gov/about/partnerships/office-of-partner-engagement/active-shooter-
resources/responding-to-an-active-shooter-crisis-situation (accessed July 21, 2019).
[42] M.J. Kinsey, E.R. Galea, P.J. Lawrence, Human Factors Associated with the Selection of
Lifts/Elevators or Stairs in Emergency and Normal Usage Conditions, Fire Technol. 48
(2012) 3–26. doi:10.1007/s10694-010-0176-7.
[43] M. Haghani, M. Sarvi, Human exit choice in crowded built environments: Investigating
underlying behavioural differences between normal egress and emergency evacuations,
Fire Saf. J. 85 (2016) 1–9. doi:10.1016/j.firesaf.2016.07.003.
[44] J. Chen, T. Shi, N. Li, Pedestrian evacuation simulation in indoor emergency situations:
Approaches, models and tools, Saf. Sci. 142 (2021) 105378.
doi:10.1016/j.ssci.2021.105378.
[45] R. Lovreglio, E. Ronchi, M.J. Kinsey, An Online Survey of Pedestrian Evacuation Model
Usage and Users, Fire Technol. 56 (2019) 1133–1153. doi:10.1007/s10694-019-00923-8.
[46] C. Arteaga, J. Park, Building design and its effect on evacuation efficiency and casualty
levels during an indoor active shooter incident, Saf. Sci. 127 (2020) 104692.
doi:10.1016/j.ssci.2020.104692.
[47] K. Hu, S. Yoon, V. Pavlovic, P. Faloutsos, M. Kapadia, Predicting Crowd Egress and
Environment Relationships to Support Building Design Optimization, Comput. Graph. 88
(2020) 83–96. doi:10.1016/j.cag.2020.03.005.
[48] X. Yang, H. Dong, Q. Wang, Y. Chen, X. Hu, Guided crowd dynamics via modified
social force model, Phys. A Stat. Mech. Its Appl. 411 (2014) 63–73.
doi:10.1016/j.physa.2014.05.068.
[49] M.L. Chu, K. Law, Computational framework incorporating human behaviors for egress
simulations, J. Comput. Civ. Eng. 27 (2013) 699–707. doi:10.1061/(ASCE)CP.1943-
5487.0000313.
[50] G. Proulx, Occupant behaviour and evacuation, Proc. 9th Int. Fire Prot. Symp. (2001)
219–232.
[51] X. Zheng, T. Zhong, M. Liu, Modeling crowd evacuation of a building based on seven
methodological approaches, Build. Environ. 44 (2009) 437–445.
doi:10.1016/j.buildenv.2008.04.002.
132

[52] A. Shipman, A. Majumdar, Fear in humans: a glimpse into the crowd-modeling
perspective, Transp. Res. Rec. 2672 (2018) 183–197. doi:10.1177/0361198118787343.
[53] A. V. Banerjee, A Simple Model of Herd Behavior, Q. J. Econ. 107 (1992) 797–817.
doi:10.2307/2118364.
[54] D. Helbing, I. Farkas, T. Vicsek, Simulating dynamical features of escape panic, Nature.
407 (2000) 487–490. doi:10.1038/35035023.
[55] M. Haghani, M. Sarvi, Following the crowd or avoiding it? Empirical investigation of
imitative behaviour in emergency escape of human crowds, Anim. Behav. 124 (2017) 47–
56. doi:10.1016/j.anbehav.2016.11.024.
[56] M. Moussaïd, M. Kapadia, T. Thrash, R.W. Sumner, M. Gross, D. Helbing, C. Hölscher,
Crowd behaviour during high-stress evacuations in an immersive virtual environment, J.
R. Soc. Interface. 13 (2016) 20160414. doi:10.1098/rsif.2016.0414.
[57] F. Alavizadeh, B. Moshiri, C. Lucas, Application of naturalistic decision making to
emergency evacuation simulations, in: Lect. Notes Comput. Sci. (Including Subser. Lect.
Notes Artif. Intell. Lect. Notes Bioinformatics), Springer Berlin Heidelberg, Berlin,
Heidelberg, 2008: pp. 14–25. doi:10.1007/978-3-540-87881-0_3.
[58] M. Haghani, M. Sarvi, Imitative (herd) behaviour in direction decision-making hinders
efficiency of crowd evacuation processes, Saf. Sci. 114 (2019) 49–60.
doi:10.1016/j.ssci.2018.12.026.
[59] M. Haghani, M. Sarvi, Stated and revealed exit choices of pedestrian crowd evacuees,
Transp. Res. Part B Methodol. 95 (2017) 238–259. doi:10.1016/j.trb.2016.10.019.
[60] N.R. Johnson, Panic and the Breakdown of Social Order: Popular Myth, Social Theory,
Empirical Evidence, Sociol. Focus. 20 (1987) 171–183.
doi:10.1080/00380237.1987.10570950.
[61] J.L. Bryan, A selected historical review of human behavior in fire, Fire Prot. Eng. 16
(2002) 4–10. http://www.frocc.org/pdf/building_eva/historical_review.pdf (accessed
December 30, 2017).
[62] C. Cocking, J. Drury, S. Reicher, The psychology of crowd behaviour in emergency
evacuations: Results from two interview studies and implications for the Fire and Rescue
Services, Irish J. Psychol. 30 (2009) 59–73. doi:10.1080/03033910.2009.10446298.
[63] I. von Sivers, A. Templeton, F. Künzner, G. Köster, J. Drury, A. Philippides, T. Neckel,
H.J. Bungartz, Modelling social identification and helping in evacuation simulation, Saf.
Sci. 89 (2016) 288–300. doi:10.1016/j.ssci.2016.07.001.
[64] J. Drury, C. Cocking, S. Reicher, A. Burton, D. Schofield, A. Hardwick, D. Graham, P.
Langston, Cooperation versus competition in a mass emergency evacuation: A new
laboratory simulation and a new theoretical model, Behav. Res. Methods. 41 (2009) 957–
133

970. doi:10.3758/BRM.41.3.957.
[65] S. Sharma, Simulation and modeling of group behavior during emergency evacuation,
Intell. Agents, 2009. IA ’09. IEEE Symp. (2009) 122–127. doi:10.1109IA.2009.4927509.
[66] N. Shiwakoti, R. Tay, P. Stasinopoulos, P.J. Woolley, Likely behaviours of passengers
under emergency evacuation in train station, Saf. Sci. 91 (2017) 40–48.
doi:10.1016/j.ssci.2016.07.017.
[67] R.L. Best, Reconstruction of a Tragedy: The Beverly Hills Supper Club Fire, Southgate,
Kentucky, May 28, 1977, National Fire Protection Association (NFPA), 1978.
[68] D. Canter, Fires and human behaviour: Emerging issues, Fire Saf. J. 3 (1980) 41–46.
[69] J.D. Averill, R.D. Peacock, E.D. Kuligowski, Analysis of the Evacuation of the World
Trade Center Towers on September 11, 2001, Fire Technol. 49 (2013) 65–81.
doi:10.1007/s10694-011-0240-y.
[70] J. Fang, S. El-Tawil, B. Aguirre, Leader-follower model for agent based simulation of
social collective behavior during egress, Saf. Sci. 83 (2016) 40–47.
doi:10.1016/j.ssci.2015.11.015.
[71] S. Pluchino, C.M.G. Tribulato, A. Caverzan, a M. Quillan, G.P. Cimellaro, a M. Asce, S.
Mahin, a M. Asce, Agent-Based Model for Pedestrians ’ Evacuation after a Blast
Integrated with a Human Behavior Model, Struct. Congr. (2015) 1506–1517.
doi:doi:10.1061/9780784479117.129.
[72] E.D. Kuligowski, S.M. V. Gwynne, The Need for Behavioral Theory in Evacuation
Modeling, Pedestr. Evacuation Dyn. 2008. (2010) 721–732. doi:10.1007/978-3-642-
04504-2_70.
[73] M. Gerges, M. Mayouf, P. Rumley, D. Moore, Human behaviour under fire situations in
high-rise residential building, Int. J. Build. Pathol. Adapt. 35 (2017) 90–106.
doi:10.1108/IJBPA-09-2016-0022.
[74] E. Ronchi, D. Nilsson, Fire evacuation in high-rise buildings: a review of human
behaviour and modelling research, Fire Sci. Rev. 2 (2013) 7. doi:10.1186/2193-0414-2-7.
[75] L. Tan, M. Hu, H. Lin, Agent-based simulation of building evacuation: Combining human
behavior with predictable spatial accessibility in a fire emergency, Inf. Sci. (Ny). 295
(2015) 53–66. doi:10.1016/j.ins.2014.09.029.
[76] M. Tancogne-Dejean, P. Laclémence, Fire risk perception and building evacuation by
vulnerable persons: Points of view of laypersons, fire victims and experts, Fire Saf. J. 80
(2016) 9–19. doi:10.1016/j.firesaf.2015.11.009.
[77] A. Rahouti, R. Lovreglio, P. Jackson, S. Datoussaid, Evacuation Data from a Hospital
Outpatient Drill : The Case Study of the North Evacuation Data from a Hospital
134

Outpatient Drill The Case Study of North Shore Hospital, in: Proc. 9th Int. Conf. Pedestr.
Evacuation Dyn., 2018: p. nd.
https://www.researchgate.net/publication/326345094_Evacuation_Data_from_a_Hospital
_Outpatient_Drill_The_Case_Study_of_the_North_Shore_Hospital.
[78] F.Z. Huo, W.G. Song, X.D. Liu, Z.G. Jiang, K.M. Liew, Investigation of human behavior
in emergent evacuation from an underground retail store, Procedia Eng. 71 (2014) 350–
356. doi:10.1016/j.proeng.2014.04.050.
[79] G. Proulx, High-rise Office Egress: The Human Factors, in: Proc. Symp. High-Rise Build.
Egr. Stairs, New York, 2007. doi:10.1039/B910216G.
[80] M. Raubal, M.J. Egenhofer, Comparing the complexity of way finding tasks in built
environments, Environ. Plan. B-Planning Des. 25 (1998) 895–913. doi:DOI
10.1068/b250895.
[81] H. Xie, L. Filippidis, E.R. Galea, D. Blackshields, P.J. Lawrence, Experimental analysis
of the effectiveness of emergency signage and its implementation in evacuation
simulation, in: Fire Mater., Wiley-Blackwell, 2012: pp. 367–382. doi:10.1002/fam.1095.
[82] M. Occhialini, G. Bernardini, F. Ferracuti, S. Iarlori, M. D’Orazio, S. Longhi, Fire exit
signs: The use of neurological activity analysis for quantitative evaluations on their
perceptiveness in a virtual environment, Fire Saf. J. 82 (2016) 63–75.
doi:10.1016/j.firesaf.2016.03.003.
[83] J. Olander, E. Ronchi, R. Lovreglio, D. Nilsson, Dissuasive exit signage for building fire
evacuation, Appl. Ergon. 59 (2017) 84–93. doi:10.1016/j.apergo.2016.08.029.
[84] M. Kobes, I. Helsloot, B. De Vries, J. Post, Exit choice, (pre-)movement time and
(pre-)evacuation behaviour in hotel fire evacuation - Behavioural analysis and validation
of the use of serious gaming in experimental research, Procedia Eng. 3 (2010) 37–51.
doi:10.1016/j.proeng.2010.07.006.
[85] J. Troncoso, D. Nilsson, E. Ronchi, Response to emergency way-finding systems by
people from different cultures, in: Hum. Behav. Fire Symp. Intersci. Commun.,
Interscience Communications, Cambridge, UK, 2015.
https://www.researchgate.net/publication/283072432_Response_to_emergency_way-
finding_systems_by_people_from_different_cultures (accessed January 1, 2020).
[86] H. Xie, E.R. Galea, P.J. Lawrence, Experimental and survey studies on the effectiveness
of dynamic signage systems, Fire Saf. Sci. 11 (2014) 1129–1143.
doi:10.3801/IAFSS.FSS.11-1129.
[87] E.R. Galea, H. Xie, S. Deere, D. Cooney, L. Filippidis, An international survey and full-
scale evacuation trial demonstrating the effectiveness of the active dynamic signage
system concept, in: Fire Mater., Wiley-Blackwell, 2017: pp. 493–513.
doi:10.1002/fam.2414.
135

[88] V.A. Oven, N. Cakici, Modelling the evacuation of a high-rise office building in Istanbul,
Fire Saf. J. 44 (2009) 1–15. doi:10.1016/j.firesaf.2008.02.005.
[89] L. Benthorn, H. Frantzich, Fire alarm in a public building: How do people evaluate
information and choose an evacuation exit?, Fire Mater. 23 (1999) 311–315.
doi:10.1002/(SICI)1099-1018(199911/12)23:6<311::AID-FAM704>3.0.CO;2-J.
[90] N.W.F. Bode, E.A. Codling, Human exit route choice in virtual crowd evacuations, Anim.
Behav. 86 (2013) 347–358. doi:10.1016/j.anbehav.2013.05.025.
[91] C. Dias, M. Sarvi, N. Shiwakoti, O. Ejtemai, M. Burd, Examining the impact of different
turning angles on the collective egress of crowds, J. Transp. Saf. Secur. 6 (2014) 167–181.
doi:10.1080/19439962.2013.831964.
[92] Z. Shahhoseini, M. Sarvi, Pedestrian crowd flows in shared spaces: Investigating the
impact of geometry based on micro and macro scale measures, Transp. Res. Part B
Methodol. 122 (2019) 57–87. doi:10.1016/j.trb.2019.01.019.
[93] E. Vilar, F. Rebelo, P. Noriega, J. Teles, C. Mayhorn, The influence of environmental
features on route selection in an emergency situation, Appl. Ergon. 44 (2013) 618–627.
doi:10.1016/j.apergo.2012.12.002.
[94] M.L. Chu, P. Parigi, K.H. Law, J.-C. Latombe, Simulating individual, group, and crowd
behaviors in building egress, Simulation. 91 (2015) 825–845.
doi:10.1177/0037549715605363.
[95] R.D. Peacock, P.A. Reneke, E.D. Kuligowski, C.R. Hagwood, Movement on Stairs
During Building Evacuations, Fire Technol. 53 (2017) 845–871. doi:10.1007/s10694-016-
0603-5.
[96] R.D. Peacock, B.L. Hoskins, E.D. Kuligowski, Overall and local movement speeds during
fire drill evacuations in buildings up to 31 stories, Saf. Sci. 50 (2012) 1655–1664.
doi:10.1016/j.ssci.2012.01.003.
[97] R.R.M. Gershon, K.A. Qureshi, R.N. Msn, M.S. Rubin, V.H. Raveis, Factors Associated
with High-Rise Evacuation : Qualitative Results from the World Trade Center Evacuation
Study, Prehosp. Disaster Med. 22 (2007) 165–173.
[98] Y. Zhang, W. Xie, S. Chen, T. Li, Experimental Study on Descent Speed on Stairs of
Individuals and Small Groups Under Different Visibility Conditions, Fire Technol. 54
(2018) 781–796. doi:10.1007/s10694-018-0710-6.
[99] E.R. Galea, G. Sharp, P.J. Lawrence, Investigating the representation of merging behavior
at the floor-stair interface in computer simulations of multi-floor building evacuations, J.
Fire Prot. Eng. 18 (2008) 291–316. doi:10.1177/1042391508095092.
[100] E. Heyes, M. Spearpoint, Lifts for evacuation-human behaviour considerations, Fire
Mater. 36 (2012) 297–308. doi:10.1002/fam.1111.
136

[101] R.F. Fahy, G. Proulx, Toward creating a database on delay times to start evacuation and
walking speeds for use in evacuation modelling, 2nd Int. Symp. Hum. Behav. Fire. (2001)
175–183. doi:10.1016/j.apergo.2012.12.002.
[102] D.A. Purser, M. Bensilum, Quantification of behaviour for engineering design standards
and escape time calculations, Saf. Sci. 38 (2001) 157–182. doi:10.1016/S0925-
7535(00)00066-7.
[103] S.M.V. Gwynne, E.D. Kuligowski, M.J. Kinsey, L.M. Hulse, Modelling and influencing
human behaviour in fire, Fire Mater. 41 (2017) 412–430. doi:10.1002/fam.2391.
[104] G. Hofinger, R. Zinke, L. Künzer, Human factors in evacuation simulation, planning, and
guidance, Transp. Res. Procedia. 2 (2014) 603–611. doi:10.1016/j.trpro.2014.09.101.
[105] P.Å. Olsson, M.A. Regan, A comparison between actual and predicted evacuation times,
Saf. Sci. 38 (2001) 139–145. doi:10.1016/S0925-7535(00)00064-3.
[106] M.T. Kinateder, E.D. Kuligowski, P.K. Reneke, R.D. Peacock, A Review of Risk
Perception in Building Fire Evacuation, (2014). doi:10.6028/NIST.TN.1840.
[107] D. Tong, D. Canter, The decision to evacuate: a study of the motivations which contribute
to evacuation in the event of fire, Fire Saf. J. 9 (1985) 257–265. doi:10.1016/0379-
7112(85)90036-0.
[108] C.M. Zhao, S.M. Lo, S.P. Zhang, M. Liu, A Post-fire Survey on the Pre-evacuation
Human Behavior, Fire Technol. 45 (2009) 71–95. doi:10.1007/s10694-007-0040-6.
[109] T. Jin, T. Yamada, Experimental Study of Human Behavior in Smoke Filled Corridors,
Fire Saf. Sci. 2 (1989) 511–519. doi:10.3801/iafss.fss.2-511.
[110] S. Cao, L. Fu, P. Wang, G. Zeng, W. Song, Experimental and modeling study on
evacuation under good and limited visibility in a supermarket, Fire Saf. J. 102 (2018) 27–
36. doi:10.1016/j.firesaf.2018.10.003.
[111] G. Proulx, R.F. Fahy, Ashrae, Human Behavior and Evacuation Movement in Smoke,
Ashrae Trans. 114 (2008) 159–165.
[112] M.K. Lindell, C.S. Prater, H.C. Wu, S.K. Huang, D.M. Johnston, J.S. Becker, H.
Shiroshita, Immediate behavioural responses to earthquakes in Christchurch, New
Zealand, and Hitachi, Japan, Disasters. 40 (2016) 85–111. doi:10.1111/disa.12133.
[113] G. Prati, E. Saccinto, L. Pietrantoni, C. Pérez-Testor, The 2012 Northern Italy
Earthquakes: Modelling human behaviour, Nat. Hazards. 69 (2013) 99–113.
doi:10.1007/s11069-013-0688-9.
[114] S. Shapira, L. Aharonson-Daniel, Y. Bar-Dayan, Anticipated behavioral response patterns
to an earthquake: The role of personal and household characteristics, risk perception,
previous experience and preparedness, Int. J. Disaster Risk Reduct. 31 (2018) 1–8.
137

doi:10.1016/j.ijdrr.2018.04.001.
[115] Z. Liu, C.C. Jacques, S. Szyniszewski, J.K. Guest, B.W. Schafer, T. Igusa, J. Mitrani-
Reiser, Agent-Based Simulation of Building Evacuation after an Earthquake : Coupling
Human Behavior with Structural Response, Nat. HAZARDS Rev. 17 (2016) 1–14.
doi:10.1061/(ASCE)NH.1527-6996.0000199.
[116] S. Gunn, P.B. Luh, X. Lu, B. Hotaling, Optimizing guidance for an active shooter event,
in: Proc. - IEEE Int. Conf. Robot. Autom., 2017: pp. 4299–4304.
doi:10.1109/ICRA.2017.7989495.
[117] Q. Liu, A social force model for the crowd evacuation in a terrorist attack, Physica A. 502
(2018) 315–330. doi:10.1016/j.physa.2018.02.136.
[118] Federal Emergency Management Agency, Reference Manual to Mitigate Potential
Terrorist Attacks Against Buildings, (2011).
[119] S. Li, J. Zhuang, S. Shen, J. Wang, Driving-forces model on individual behavior in
scenarios considering moving threat agents, Phys. A Stat. Mech. Its Appl. 481 (2017)
127–140. doi:10.1016/j.physa.2017.04.011.
[120] S. Li, J. Zhuang, S. Shen, A three-stage evacuation decision-making and behavior model
for the onset of an attack, Transp. Res. Part C Emerg. Technol. 79 (2017) 119–135.
doi:10.1016/j.trc.2017.03.008.
[121] J. Chen, J. Wang, B. Wang, R. Liu, Q. Wang, An experimental study of visibility effect on
evacuation speed on stairs, Fire Saf. J. 96 (2018) 189–202.
doi:10.1016/j.firesaf.2017.11.010.
[122] R.Y. Guo, H.J. Huang, S.C. Wong, Route choice in pedestrian evacuation under
conditions of good and zero visibility: Experimental and simulation results, Transp. Res.
Part B Methodol. 46 (2012) 669–686. doi:10.1016/j.trb.2012.01.002.
[123] T. Gärling, E. Lindberg, T. Mäntylä, Orientation in buildings: Effects of familiarity, visual
access, and orientation aids, J. Appl. Psychol. 68 (1983) 177–186. doi:10.1037/0021-
9010.68.1.177.
[124] A. Seidel, Way-finding in public spaces: The Dallas/Fort Worth, USA airport, in: 20th Int.
Congr. Appl. Psychol., 1982.
[125] J.R. Carpman, M.A. Grant, D.A. Simmons, Hospital design and wayfinding: A Video
Simulation Study, Environ. Behav. 17 (1985) 296–314. doi:10.1177/0013916585173002.
[126] A. Sixsmithm, J. Sixsmith, D. Canter, When is a door not a door? A study of evacuation
route identification in a large shopping mall, Saf. Built Environ. (1988) 62–74.
[127] G. Weisman, Orientation, Path Finding, and Architectural Legibility: a Review and
Theoretical Integration, in: Proc. Int. Conf. Build. Use Saf. Technol., 1985: pp. 9–15.
138

[128] Interagency Security Committee, Planning and Response to an Active Shooter: An
Interagency Security Committee Policy and Best Practices Guide, (2015).
[129] InterAgency Board, Active Shooter/Hostile Event (ASHE) Guide, 2016.
[130] E.D. Kuligowski, Guidance Document: Emergency Communication Strategies for
Buildings, 2014.
[131] C. Cho, J.W. Park, S. Sakhakarmi, Emergency response: Effect of human detection
resolution on risks during indoor mass shooting events, Saf. Sci. 114 (2019) 160–170.
doi:10.1016/j.ssci.2019.01.021.
[132] J.Y. Lee, J. Eric Dietz, K. Ostrowski, Agent-based modeling for casualty rate assessment
of large event active shooter incidents, Proc. - Winter Simul. Conf. (2019) 2737–2746.
doi:10.1109/WSC.2018.8632535.
[133] S.F. Greenberg, Active Shooters on College Campuses: Conflicting Advice, Roles of the
Individual and First Responder, and the Need to Maintain Perspective, Disaster Med.
Public Health Prep. 1 (2007) S57–S61. doi:10.1097/dmp.0b013e318149f492.
[134] D.N. Sattler, K. Larpenteur, G. Shipley, Active Shooter on Campus: Evaluating Text and
E-mail Warning Message Effectiveness, J. Homel. Secur. Emerg. Manag. 8 (2011).
doi:10.2202/1547-7355.1826.
[135] J.A. Fox, J. Savage, Mass Murder Goes to College: An Examination of Changes on
College Campuses Following Virginia Tech, Am. Behav. Sci. 52 (2009) 1465–1485.
doi:10.1177/0002764209332558.
[136] Federal Emergency Management Agency, Primer to Design Safe School Projects in Case
of Terrorist Attacks, (2003).
[137] Federal Emergency Management Agency, Primer for Design of Commercial Buildings to
Mitigate Terrorist Attacks, (2003).
[138] Department of Defense, DoD Minimum Antiterrorism Standards for Buildings, (2013).
[139] G. Santos, B.E. Aguirre, A Critical Review Of Emergency Evacuation Simulation Models,
in: NIST Work. Build. Occupant Mov. Dur. Fire Emergencies, 2004.
[140] B. Levin, EXITT- A Simulation Model Of Occupant Decisions And Actions In
Residential Fires, Fire Saf. Sci. 2 (1989) 561–570. doi:10.3801/IAFSS.FSS.2-561.
[141] S. Wolfram, Statistical mechanics of cellular automata, Rev. Mod. Phys. 55 (1983) 601–
644. doi:10.1103/RevModPhys.55.601.
[142] Y. Song, K. Xie, W. Su, Mechanism and strategies of post-earthquake evacuation based
on cellular automata model, Int. J. Disaster Risk Reduct. 34 (2019) 220–231.
doi:10.1016/j.ijdrr.2018.11.020.
139

[143] M. MacDonald, STPES-simulating pedestrian dynamics, (2022).
https://www.steps.mottmac.com/steps-dynamics.
[144] D. Helbing, P. Molnár, Social force model for pedestrian dynamics, Phys. Rev. E. 51
(1995) 4282–4286. doi:10.1103/PhysRevE.51.4282.
[145] H. Zhang, H. Liu, X. Qin, B. Liu, Modified two-layer social force model for emergency
earthquake evacuation, Phys. A Stat. Mech. Its Appl. 492 (2018) 1107–1119.
doi:10.1016/J.PHYSA.2017.11.041.
[146] T. Korhonen, S. Hostikka, Fire Dynamics Simulator with Evacuation: FDS+Evac,
Technical Reference and User’s Guide, 2010.
[147] P. Group, PTV Viswalk, (2022). https://www.ptvgroup.com/en/solutions/products/ptv-
viswalk/.
[148] Simwalk, SIMWALK – Crowd analysis for architecture and urban planning, (2022).
https://simwalk.com/simwalk_pro/index.html.
[149] E. Bonabeau, Agent-based modeling: Methods and techniques for simulating human
systems, Proc. Natl. Acad. Sci. U. S. A. 99 (2002) 7280–7287.
doi:10.1073/pnas.082080899.
[150] X. Pan, C.S. Han, K. Dauber, K.H. Law, Human and social behavior in computational
modeling and analysis of egress, Autom. Constr. 15 (2006) 448–461.
doi:10.1016/j.autcon.2005.06.006.
[151] T. Engineering, Pathfinder, (2022). https://www.thunderheadeng.com/pathfinder/.
[152] Oasys, Crowd simulation software: MassMotion, (2022). https://www.oasys-
software.com/products/pedestrian-simulation/massmotion/.
[153] T.A. Company, AnyLogic: Simulation Modeling Software Tools & Solutions for
Business, (2022). https://www.anylogic.com/.
[154] M. Busogi, D. Shin, H. Ryu, Y.G. Oh, N. Kim, Weighted affordance-based agent
modeling and simulation in emergency evacuation, Saf. Sci. 96 (2017) 209–227.
doi:10.1016/j.ssci.2017.04.005.
[155] K. Rendón Rozo, J. Arellana, A. Santander-Mercado, M. Jubiz-Diaz, Modelling building
emergency evacuation plans considering the dynamic behaviour of pedestrians using
agent-based simulation, Saf. Sci. 113 (2019) 276–284. doi:10.1016/j.ssci.2018.11.028.
[156] M. Haghani, M. Sarvi, Simulating dynamics of adaptive exit-choice changing in crowd
evacuations: Model implementation and behavioural interpretations, Transp. Res. Part C
Emerg. Technol. 103 (2019) 56–82. doi:10.1016/j.trc.2019.04.009.
[157] X. Zhao, R. Lovreglio, D. Nilsson, Modelling and interpreting pre-evacuation decision-
140

making using machine learning, Autom. Constr. 113 (2020) 103140.
doi:10.1016/j.autcon.2020.103140.
[158] K. Wang, X. Shi, A.P.X. Goh, S. Qian, A machine learning based study on pedestrian
movement dynamics under emergency evacuation, Fire Saf. J. 106 (2019) 163–176.
doi:10.1016/j.firesaf.2019.04.008.
[159] Q. Sun, Y. Turkan, A BIM-based simulation framework for fire safety management and
investigation of the critical factors affecting human evacuation performance, Adv. Eng.
Informatics. 44 (2020) 101093. doi:10.1016/j.aei.2020.101093.
[160] Z. Fang, W. Song, J. Zhang, H. Wu, Experiment and modeling of exit-selecting behaviors
during a building evacuation, Phys. A Stat. Mech. Its Appl. 389 (2010) 815–824.
doi:10.1016/j.physa.2009.10.019.
[161] W. Li, J. Gong, P. Yu, S. Shen, R. Li, Q. Duan, Simulation and analysis of congestion risk
during escalator transfers using a modified social force model, Phys. A Stat. Mech. Its
Appl. 420 (2015) 28–40. doi:10.1016/j.physa.2014.10.044.
[162] S.M. V. Gwynne, Translating Behavioral Theory of Human Response into Modeling
Practice, Natl. Inst. Stand. Technol. (2012) 69. http://dx.doi.org/10.6028/NIST.GCR.972.
[163] A. Wong, Hong Kong Subway Fire Injures at Least 17 People, (2017).
https://www.nytimes.com/2017/02/10/world/asia/hong-kong-subway-fire.html (accessed
August 28, 2019).
[164] National Standardization Management Committee, Code for safety metro evacuation
(GB/T 33668—2017), 2017.
[165] G. Hoftede, G.J. Hofstede, M. Minkov, Cultures and Organizations: Software of the Mind,
Mcgraw-Hill, New York, 2010.
[166] Londonist, London Underground Vs Beijing Subway, (2017).
https://londonist.com/2015/10/london-underground-vs-beijing-subway (accessed
September 10, 2019).
[167] F. Anderton, Comparing LA’s transit system to London and New York, (2017).
https://www.kcrw.com/culture/shows/design-and-architecture/comparing-las-transit-
system-to-london-and-new-york (accessed September 10, 2019).
[168] J. Lin, L. Cao, N. Li, Assessing the influence of repeated exposures and mental stress on
human wayfinding performance in indoor environments using virtual reality technology,
Adv. Eng. Informatics. 39 (2019) 53–61. doi:10.1016/j.aei.2018.11.007.
[169] L. Qiu, X. Zheng, Y. Wang, Revision of the Positive Affect and Negative Affect Scale,
Chinese J. Appl. Psychol. 14 (2008) 249–254.
[170] D. Watson, L. a. Clark, A. Tellegan, Development and validation of brief measures of
141

positive and negative affect: The PANAS scales Worksheet 3.1 PANAS Schedule
Questionnaire, J. Pers. Soc. Psychol. 54 (1988) 1063–1070.
[171] R.S. Kennedy, N.E. Lane, S. Kevin, M.G. Lilienthal, Simulator Sickness Questionnaire :
An Enhanced Method for Quantifying Simulator Sickness, Int. J. Aviat. Psychol. 3 (1993)
203–220. doi:10.1207/s15327108ijap0303.
[172] M. Hegarty, A.E. Richardson, D.R. Montello, K. Lovelace, I. Subbiah, Development of a
self-report measure of environmental spatial ability, Intelligence. 30 (2002) 425–447.
[173] B.G. Witmer, M.J. Singer, Measuring presence in virtual environments: A presence
questionnaire, Presence Teleoperators Virtual Environ. 7 (1998) 225–240.
doi:10.1162/105474698565686.
[174] C.A. Lawton, Gender differences in way-finding strategies: Relationship to spatial ability
and spatial anxiety, Sex Roles. 30 (1994) 765–779.
[175] M. Usoh, E. Catena, S. Arman, M. Slater, Using presence questionnaires in reality,
Presence Teleoperators Virtual Environ. 9 (2000) 497–503.
doi:10.1162/105474600566989.
[176] D. Purser, ASET and RSET: Addressing some issues in relation to occupant behaviour
and tenability, in: Fire Saf. Sci., 2003: pp. 91–102. doi:10.3801/IAFSS.FSS.7-91.
[177] M.T. Kinateder, E. Ronchi, D. Gromer, M. Müller, M. Jost, M. Nehfischer, A.
Mühlberger, P. Pauli, Social influence on route choice in a virtual reality tunnel fire,
Transp. Res. Part F Traffic Psychol. Behav. 26 (2014) 116–125.
doi:10.1016/j.trf.2014.06.003.
[178] M.J. Hurley, D. Gottuk, J.R. Hall, K. Harada, E. Kuligowski, M. Puchovsky, J. Torero,
J.M. Watts, C. Wieczorek, eds., SFPE Handbook of Fire Protection Engineering, Springer
New York, New York, NY, 2016. doi:10.1007/978-1-4939-2565-0.
[179] G. Proulx, Evacuation time and movement in apartment buildings, Fire Saf. J. 24 (1995)
229–246. doi:10.1016/0379-7112(95)00023-M.
[180] K.L. Yoon, D.N. Vidaurri, J. Joormann, R. De Raedt, Social anxiety and narrowed
attentional breadth toward faces, Emotion. 15 (2015) 682–686. doi:10.1037/emo0000070.
[181] X. Lu, P.B. Luh, A. Tucker, T. Gifford, R.S. Astur, N. Olderman, Impacts of Anxiety in
Building Fire and Smoke Evacuation: Modeling and Validation, IEEE Robot. Autom.
Lett. 2 (2017) 255–260. doi:10.1109/LRA.2016.2579744.
[182] P.G. Wood, The Behaviour of People in Fires, Fire Saf. Sci. 953 (1972).
http://iafss.org/publications/frn/953/-1 (accessed February 21, 2018).
[183] M. Haghani, M. Sarvi, Pedestrian crowd tactical-level decision making during emergency
evacuations, J. Adv. Transp. 50 (2016) 1870–1895. doi:10.1002/atr.1434.
142

[184] R. Lovreglio, A. Fonzone, L. Dell’Olio, D. Borri, A. Ibeas, The Role of Herding
Behaviour in Exit Choice During Evacuation, in: Procedia - Soc. Behav. Sci., 2014: pp.
390–399.
[185] I. Wagner, W. Schnotz, Learning from static and dynamic visualizations: What kind of
questions should we ask?, in: Learn. from Dyn. Vis., Springer, 2017: pp. 69–91.
[186] G. Proulx, A stress model for people facing a fire, J. Environ. Psychol. 13 (1993) 137–
147. doi:10.1016/S0272-4944(05)80146-X.
[187] A. Tucker, K.L. Marsh, T. Gifford, X. Lu, P.B. Luh, R.S. Astur, The effects of
information and hazard on evacuee behavior in virtual reality, Fire Saf. J. 99 (2018) 1–11.
doi:10.1016/j.firesaf.2018.04.011.
[188] D. Stewart, P. Shamdasani, D. Rook, Analyzing Focus Group Data, Sage publications,
2007. doi:10.4135/9781412991841.
[189] S.M. Van Es, E.M. Le Coq, A.I. Brouwer, I. Mesters, A.F. Nagelkerke, V.T. Colland,
Adherence-related behavior in adolescents with asthma: Results from focus group
interviews, J. Asthma. 35 (1998) 637–646. doi:10.3109/02770909809048966.
[190] L.C. Lederman, Assessing educational effectiveness: The focus group interview as a
technique for data collection, Commun. Educ. 39 (1990) 117–127.
doi:10.1080/03634529009378794.
[191] C.A. Richardson, F. Rabiee, Question of access : An exploration of the factors that
influence the health of young use of health care services, Heal. Educ. J. 60 (2001) 3–16.
[192] D.W. Stewart, P.N. Shamdasani, Focus Groups: Theory and Practice., Sage publications,
2014.
[193] A. Howard, K. Agllias, M. Bevis, T. Blakemore, “They’ll tell us when to evacuate”: The
experiences and expectations of disaster-related communication in vulnerable groups, Int.
J. Disaster Risk Reduct. 22 (2017) 139–146. doi:10.1016/j.ijdrr.2017.03.002.
[194] C.A. Tuttas, Lessons learned using web conference technology for online focus group
interviews, Qual. Health Res. 25 (2015) 122–133. doi:10.1177/1049732314549602.
[195] M.Q. Patton, Qualitative evaluation and research methods, Sage publications, 1990.
[196] H.W. Asfaw, Sandy Lake First Nation, T.K. McGee, A.C. Christianson, Evacuation
preparedness and the challenges of emergency evacuation in Indigenous communities in
Canada: The case of Sandy Lake First Nation, Northern Ontario, Int. J. Disaster Risk
Reduct. 34 (2019) 55–63. doi:10.1016/j.ijdrr.2018.11.005.
[197] R.W. Mcdaniel, C.A. Bach, Focus Groups: A Data-Gathering Strategy for Nursing
Research, Nurs. Sci. Q. 7 (1994) 4–5. doi:10.1177/089431849400700103.
143

[198] R.A. Krueger, M.A. Casey, Focus groups: a practical guide for applied research, Sage
publications, 2014.
[199] S. Elo, H. Kyngäs, The qualitative content analysis process, J. Adv. Nurs. 62 (2008) 107–
115. doi:10.1111/j.1365-2648.2007.04569.x.
[200] A.J. Onwuegbuzie, W.B. Dickinson, N.L. Leech, A.G. Zoran, A Qualitative Framework
for Collecting and Analyzing Data in Focus Group Research, Int. J. Qual. Methods. 8
(2009) 1–21. doi:10.1177/160940690900800301.
[201] P.S. Kidd, M.B. Parshall, Getting the focus and the group: enhancing analytical rigor in
focus group research, Qual. Health Res. 10 (2000) 293–308.
[202] J. Corbin, A. Strauss, Basics of qualitative research: Techniques and procedures for
developing grounded theory, Sage publications, 2014.
[203] L.A. Addington, Cops and cameras: Public school security as a policy response to
columbine, Am. Behav. Sci. 52 (2009) 1426–1446. doi:10.1177/0002764209332556.
[204] D.A. Drysdale, W. Modzeleski, A.B. Simons, Campus Attacks: Targeted Violence
Affecting Institutions of Higher Education, 2010.
[205] E. Kuligowski, Predicting Human Behavior During Fires, Fire Technol. 49 (2013) 101–
120. doi:10.1007/s10694-011-0245-6.
[206] R.W. Perry, M.K. Lindell, Understanding Citizen Response to Disasters with Implications
for Terrorism, J. Contingencies Cris. Manag. 11 (2003) 49–60. doi:10.1111/1468-
5973.1102001.
[207] S. Cao, X. Liu, M. Chraibi, P. Zhang, W. Song, Characteristics of pedestrian’s evacuation
in a room under invisible conditions, Int. J. Disaster Risk Reduct. 41 (2019) 101295.
doi:10.1016/j.ijdrr.2019.101295.
[208] R. Zhu, J. Lin, B. Becerik-Gerber, N. Li, Influence of architectural visual access on
emergency wayfinding: A cross-cultural study in China, United Kingdom and United
States, Fire Saf. J. 113 (2020) 102963. doi:10.1016/j.firesaf.2020.102963.
[209] R. Nagai, T. Nagatani, M. Isobe, T. Adachi, Effect of exit configuration on evacuation of
a room without visibility, Phys. A Stat. Mech. Its Appl. 343 (2004) 712–724.
[210] Department of Homeland Security, Active Shooter Preparedness, (2019).
https://www.dhs.gov/cisa/active-shooter-preparedness (accessed August 21, 2019).
[211] J.P. O’Neill, J.J. Miller, J.R. Waters, Active Shooter Recommendations and Analysis for
Risk Mitigation, 2016.
[212] E.A. Skryabina, N. Betts, G. Reedy, P. Riley, R. Amlôt, The role of emergency
preparedness exercises in the response to a mass casualty terrorist incident: A mixed
144

methods study, Int. J. Disaster Risk Reduct. 46 (2020) 101503.
doi:10.1016/j.ijdrr.2020.101503.
[213] M. Cha, S. Han, J. Lee, B. Choi, A virtual reality based fire training simulator integrated
with fire dynamics data, Fire Saf. J. 50 (2012) 12–24. doi:10.1016/j.firesaf.2012.01.004.
[214] R. Lovreglio, V. Gonzalez, Z. Feng, R. Amor, M. Spearpoint, J. Thomas, M. Trotter, R.
Sacks, Prototyping virtual reality serious games for building earthquake preparedness: The
Auckland City Hospital case study, Adv. Eng. Informatics. 38 (2018) 670–682.
doi:10.1016/j.aei.2018.08.018.
[215] T. Griffith, J. Ablanedo, T. Dwyer, Leveraging a virtual environment to prepare for school
shootings, in: Int. Conf. Virtual, Augment. Mix. Real., Springer, Cham, 2017: pp. 325–
338. doi:10.1007/978-3-319-57987-0_26.
[216] J. Jones, R. Kue, P. Mitchell, S.G. Eblan, K.S. Dyer, Emergency medical services
response to active shooter incidents: Provider comfort level and attitudes before and after
participation in a focused response training program, Prehosp. Disaster Med. 29 (2014)
350–357. doi:10.1017/S1049023X14000648.
[217] M.C. Weber, S.E. Schulenberg, E.C. Lair, University employees’ preparedness for natural
hazards and incidents of mass violence: An application of the extended parallel process
model, Int. J. Disaster Risk Reduct. 31 (2018) 1082–1091.
doi:10.1016/j.ijdrr.2018.03.032.
[218] G. Proulx, J. Pineau, Differences in the Evacuation Behaviour of Office and Apartment
Building Occupants, Proc. Hum. Factors Ergon. Soc. Annu. Meet. 40 (1996) 825–829.
doi:10.1177/154193129604001607.
[219] R. Zhu, G.M. Lucas, B. Becerik-Gerber, E.G. Southers, Building preparedness in response
to active shooter incidents: Results of focus group interviews, Int. J. Disaster Risk Reduct.
48 (2020) 101617. doi:10.1016/j.ijdrr.2020.101617.
[220] R. Zhu, B. Becerik-Gerber, G. Lucas, E. Southers, D. V. Pynadath, Information
Requirements for Virtual Environments to Study Human-Building Interactions during
Active Shooter Incidents, in: Comput. Civ. Eng., Atlanta, GA, 2019: pp. 188–195.
[221] M. Awada, R. Zhu, B. Becerik-Gerber, G. Lucas, E. Southers, An integrated emotional
and physiological assessment for VR-based active shooter incident experiments, Adv.
Eng. Informatics. 47 (2021) 101227. doi:10.1016/j.aei.2020.101227.
[222] R. Lovreglio, A. Fonzone, L. dell’Olio, A mixed logit model for predicting exit choice
during building evacuations, Transp. Res. Part A Policy Pract. 92 (2016) 59–75.
doi:10.1016/j.tra.2016.06.018.
[223] J. Schildkraut, A.B. Nickerson, Ready to Respond: Effects of Lockdown Drills and
Training on School Emergency Preparedness, Vict. Offenders. 15 (2020) 619–638.
doi:10.1080/15564886.2020.1749199.
145

[224] C.L. Jonson, M.M. Moon, J.A. Hendry, One Size Does Not Fit All: Traditional Lockdown
Versus Multioption Responses to School Shootings, J. Sch. Violence. 19 (2020) 154–166.
doi:10.1080/15388220.2018.1553719.
[225] M. Haghani, Optimising crowd evacuations: Mathematical, architectural and behavioural
approaches, Saf. Sci. 128 (2020) 104745. doi:10.1016/j.ssci.2020.104745.
[226] H. Gao, B. Medjdoub, H. Luo, H. Zhong, B. Zhong, D. Sheng, Building evacuation time
optimization using constraint-based design approach, Sustain. Cities Soc. 52 (2020)
101839. doi:10.1016/j.scs.2019.101839.
[227] J. Lin, L. Cao, N. Li, How the completeness of spatial knowledge influences the
evacuation behavior of passengers in metro stations: A VR-based experimental study,
Autom. Constr. 113 (2020) 103136. doi:10.1016/j.autcon.2020.103136.
[228] J. Lin, R. Zhu, N. Li, B. Becerik-Gerber, Do people follow the crowd in building
emergency evacuation? A cross-cultural immersive virtual reality-based study, Adv. Eng.
Informatics. 43 (2020) 101040. doi:10.1016/j.aei.2020.101040.
[229] Z. Feng, V.A. González, R. Amor, R. Lovreglio, G. Cabrera-Guerrero, Immersive virtual
reality serious games for evacuation training and research: A systematic literature review,
Comput. Educ. 127 (2018) 252–266. doi:https://doi.org/10.1016/j.compedu.2018.09.002.
[230] P.F. Johnson, C.E. Johnson, C. Sutherland, Stay or Go? Human Behavior and Decision
Making in Bushfires and Other Emergencies, Fire Technol. 48 (2012) 137–153.
doi:10.1007/s10694-011-0213-1.
[231] L. Faul, D. Stjepanovic, J.M. Stivers, G.W. Stewart, J.L. Graner, R.A. Morey, K.S. LaBar,
Proximal threats promote enhanced acquisition and persistence of reactive fear-learning
circuits, Proc. Natl. Acad. Sci. U. S. A. 117 (2020) 16678–16689.
doi:10.1073/pnas.2004258117.
[232] H. Zou, N. Li, L. Cao, Emotional Response–Based Approach for Assessing the Sense of
Presence of Subjects in Virtual Building Evacuation Studies, J. Comput. Civ. Eng. 31
(2017) 1–10. doi:10.1061/(ASCE)CP.1943-5487.0000679.
[233] X. Zhao, X. Yan, A. Yu, P. Van Hentenryck, Prediction and behavioral analysis of travel
mode choice: A comparison of machine learning and logit models, Travel Behav. Soc. 20
(2020) 22–35. doi:10.1016/j.tbs.2020.02.003.
[234] C. Xie, J. Lu, E. Parkany, Work Travel Mode Choice Modeling with Data Mining:
Decision Trees and Neural Networks, Transp. Res. Rec. (2003) 50–61. doi:10.3141/1854-
06.
[235] J. Hagenauer, M. Helbich, A comparative study of machine learning classifiers for
modeling travel mode choice, Expert Syst. Appl. 78 (2017) 273–282.
doi:10.1016/j.eswa.2017.01.057.
146

[236] M. Liu, Y. Jiang, Measuring accessibility of urban scales: A trip-based interaction
potential model, Adv. Eng. Informatics. 48 (2021) 101293.
doi:10.1016/j.aei.2021.101293.
[237] J.C. Milton, V.N. Shankar, F.L. Mannering, Highway accident severities and the mixed
logit model: An exploratory empirical analysis, Accid. Anal. Prev. 40 (2008) 260–266.
doi:10.1016/j.aap.2007.06.006.
[238] C.F. Manski, The Structure of Random Utility Models, Theory Decis. 8 (1977) 229–254.
[239] A. Motamedi, Z. Wang, N. Yabuki, T. Fukuda, T. Michikawa, Signage visibility analysis
and optimization system using BIM-enabled virtual reality (VR) environments, Adv. Eng.
Informatics. 32 (2017) 248–262. doi:10.1016/j.aei.2017.03.005.
[240] M. Moussaïd, D. Helbing, S. Garnier, A. Johansson, M. Combe, G. Theraulaz,
Experimental study of the behavioural mechanisms underlying self-organization in human
crowds, Proc. R. Soc. B Biol. Sci. 276 (2009) 2755–2762. doi:10.1098/rspb.2009.0405.
[241] D. McFadden, Conditional logit analysis of qualitative choice behavior, Front. Econom.
(1974) 105–142.
[242] L. Cheng, X. Chen, J. De Vos, X. Lai, F. Witlox, Applying a random forest method
approach to model travel mode choice behavior, Travel Behav. Soc. 14 (2019) 1–10.
doi:10.1016/j.tbs.2018.09.002.
[243] D. Brownstone, K. Train, Forecasting new product penetration with flexible substitution
patterns, J. Econom. 89 (1998) 109–129. doi:10.1016/S0304-4076(98)00057-8.
[244] F. Wang, C.L. Ross, Machine Learning Travel Mode Choices: Comparing the
Performance of an Extreme Gradient Boosting Model with a Multinomial Logit Model,
Transp. Res. Rec. 2672 (2018) 35–45. doi:10.1177/0361198118773556.
[245] A. Natapov, A. Parush, L. Laufer, D. Fisher-Gewirtzman, Architectural features and
indoor evacuation wayfinding: The starting point matters, Saf. Sci. 145 (2022) 105483.
doi:10.1016/j.ssci.2021.105483.
[246] S. Hassanpour, A.A. Rassafi, V.A. González, J. Liu, A hierarchical agent-based approach
to simulate a dynamic decision-making process of evacuees using reinforcement learning,
J. Choice Model. 39 (2021). doi:10.1016/j.jocm.2021.100288.
[247] Y. Zhang, Z. Chai, G. Lykotrafitis, Deep reinforcement learning with a particle dynamics
environment applied to emergency evacuation of a room with obstacles, Phys. A Stat.
Mech. Its Appl. 571 (2021) 125845. doi:10.1016/j.physa.2021.125845.
[248] X. Lu, R. Astur, T. Gifford, Effects of gunfire location information and guidance on
improving survival in virtual mass shooting events, Int. J. Disaster Risk Reduct. 64
(2021). doi:10.1016/j.ijdrr.2021.102505.

Abstract (if available)

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Understanding human-building interactions through perceptual decision-making processes

PDF

Enabling human-building communication to promote pro-environmental behavior in office buildings

PDF

Towards health-conscious spaces: building for human well-being and performance

PDF

In-situ quality assessment of scan data for as-built models using building-specific geometric features

PDF

Impact of virtual reality (VR)-based training on construction robotics remote-operation

PDF

A radio frequency based indoor localization framework for supporting building emergency response operations

PDF

Behavioral form finding using multi-agent systems: a computational methodology for combining generative design with environmental and structural analysis in architectural design

PDF

Semantic modeling of outdoor scenes for the creation of virtual environments and simulations

PDF

Point cloud data fusion of RGB and thermal information for advanced building envelope modeling in support of energy audits for large districts

PDF

User-centric smart sensing for non-intrusive electricity consumption disaggregation in buildings

PDF

A framework for comprehensive assessment of resilience and other dimensions of asset management in metropolis-scale transport systems

PDF

Healthy mobility: untangling the relationships between the built environment, travel behavior, and environmental health

PDF

Behavioral signal processing: computational approaches for modeling and quantifying interaction dynamics in dyadic human interactions

PDF

Modeling human regulation of momentum while interacting with the environment

PDF

Modeling human bounded rationality in opportunistic security games

PDF

Towards social virtual listeners: computational models of human nonverbal behaviors

PDF

Human-building integration: Investigation of human eye pupil sizes as a measure of visual sensation in the workstation environment

PDF

Coordinating social communication in human-robot task collaborations

PDF

Situated proxemics and multimodal communication: space, speech, and gesture in human-robot interaction

PDF

Computational foundations for mixed-motive human-machine dialogue

Asset Metadata

Creator Zhu, Runhe (author)

Core Title Understanding human-building-emergency interactions in the built environment

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Civil Engineering

Degree Conferral Date 2022-12

Publication Date 03/07/2024

Defense Date 08/24/2022

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag built environment,emergency,Human behavior,interaction,OAI-PMH Harvest,Safety

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Becerik-Gerber, Burcin (committee chair), Lucas, Gale (committee member), Meshkati, Najmedin (committee member), Soibelman, Lucio (committee member), Southers, Erroll (committee member)

Creator Email runhe.zhu@gmail.com,runhezhu@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC111764650

Unique identifier UC111764650

Legacy Identifier etd-ZhuRunhe-11184

Document Type Dissertation

Format application/pdf (imt)

Rights Zhu, Runhe

Type texts

Source 20220908-usctheses-batch-978 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu