Computer Science Technical Report Archive

USC Computer Science Technical Reports, no. 880 (2006)

(USC DC Other)

USC Computer Science Technical Reports, no. 880 (2006)

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content A Vision System for Personal Service Robots:
Resilient Detection and Tracking of People
Alexandre R.J. Fran¸ cois and G´ erard G. Medioni
Institute for Robotics and Intelligent Systems
University of Southern California
afrancoi,medioni@usc.edu
June 2006
Abstract
This report describes the design and implementation of a computer vision system that performs real-
time people detection and tracking from stereo color video. The system, developed in the context of a
project that aims to empower personal robots with advanced vision capabilities, illustrates a systematic
approach to combining efficient but brittle algorithms into robust systems. The system’s tracking core
builds upon established monocular algorithms for face detection and color-based tracking to perform
stereo multi-target people (head) tracking. These algorithms interact in the dynamic system according
to a general tracking pattern whose core is a feedback loop. Systems designed according to this pattern
exhibit resilience properties, such as bootstrapping capability, robustness to input data noise, adaptabil-
ity to changing environment. As a stand-alone integrated demonstration, the system incorporates video
capture, processing and result visualization for evaluation purposes. It was designed using the Software
Architecture for Immersipresence model. Parallel implementations on both Windows and Linux plat-
forms, using the open source MFSM architectural middleware and the Intel OpenCV library, reliably
detect and track people at interactive rates.
1 Introduction
This report describes the design and implementation of Stevi v.1, a demonstration computer vision sub-
system that performs real-time people detection and tracking from stereo color video. Stevi v.1, developed
in the context of a project that aims to empower personal service robots with advanced vision capabilities,
illustrates a systematic approach to combining efficient algorithms into robust systems.
1.1 Motivation
Personalservicerobots,whosepurposeisto“educate,entertain,orassist,orprotectinthehome”(asdefined
by Robotic Trends magazine) are attracting significant interest in both the research and commercial commu-
nities. Suchrobotscanbeespeciallyusefulforexampleinassistingpeoplewithdisabilitiesandinaddressing
the needs of the aging population in industrial societies. A service robot must autonomously interact with
its environment in a naturalistic manner [16], and in particular be able to interact and communicate with
ordinary people in a natural and social way [9, 5, 8]. It is generally recognized that robots designed to inter-
actwithhumans, inenvironmentsdesignedbyandforhumans, musthavehuman-orientedperception[9]. In
order fora personal servicerobotto function, it musthavesomeawarenessof itsenvironment. Inparticular,
it is critical for the robot to know whether, when and where people are present. Stevi v.1 addresses this
particular capability.
Visionforrobotsencompassesmostofthedifficultchallengeswhileaddingsomespecificones. Thevision
subsystem must operate in concert with the overall (dynamic) robotic system, at rates that are consistent
with human expectations. Furthermore, mobile robots must operate robustly in unstructured environments.
Consequently, efficient but brittle vision algorithms cannot individually provide satisfactory solutions.
1
USC/CS/06-880 c ° 2006 arjf+ggm 2
Figure 1: Composite images of tracking data overlayed on left and right input color images. Note the
head area size consistent with the person’s distance to the cameras (the further away, the smaller) and the
consistent labels in the left and right views.
1.2 Overview of Approach
The key to building robust and efficient computer vision systems consists in designing complex, integrated
systems whose overall properties go beyond the sum of that of their individual algorithms. Stevi v.1’s
tracking core builds upon established monocular algorithms for face detection and color-based tracking to
perform stereo multi-target people (head) tracking. A classifier-based face detection algorithm based on [21]
performs initial target acquisition. Each detected face area instantiates a target candidate to which is
attached a color model (histogram) initialized from the face area in each image of the stereo pair. A mean
shift-based algorithm [6] tracks the targets (or rather their color models) that have sufficient combined
support in both images.
Interaction between data and processing elements implementing these algorithms in the system are cap-
turedinthesystem’sarchitecturalspecification,andcanbeabstractedintoarchitecturalpatterns. Steviv.1’s
tracking core is a feedback loop. If all tracking algorithms are feedback loops by nature, the static formu-
lation and sequential implementation of tracking systems often hide or abstract their dynamic properties.
For example, the fundamental nature of tracking as feedback is implicit in iterative methods such as mean
shift-based tracking [11]. Explicitly modeling the dynamic aspect of tracking allows to address dynamic
properties such as bootstrapping, robustness to input data noise, adaptability to changing environment.
Making explicit the dynamic nature of tracking systems also allows appropriate and efficient application of
traditional static algorithms to dynamic data.
Tracking as feedback is not a new concept. For example, Jebara and Pentland describe a 3D adaptive
feedback tracking system for faces [15]. The system switches between two modes: robust face detection
and fast tracking. It is designed as two feedback loops operating at very different rates to accommodate
the complexity of the algorithms involved and the computing power available on contemporary hardware.
Stevi v.1 illustrates a general feed-back pattern that presents an integrated approach to detection and
tracking. This pattern is not specific of the particular algorithms used. It is, by design, open to the parallel
operation of multiple algorithms (e.g. for detection) and provides a formal framework for implementing
fusion algorithms.
Specific integration of stereo processing and color-based tracking has been explored before. For example,
Morenoetal. describeareal-timesystemthatperforms3Dheadtracking[20]. Depthinformationcomputed
from stereo scales the sampling area for mean shift tracking. Stevi v.1’s system never computes pixel-level
depth information from stereo, but rather uses stereo information to estimate the depth of higher-level
objects (e.g. faces). This efficient use of stereo information increases the robustness of the system through
redundancy and enhanced adaptivity.
As a stand-alone integrated demonstration, the Stevi v.1 system incorporates video capture, processing
and result visualization for evaluation purposes. Figure 1 shows composite images of tracking information
overlayed on the input left and right images. The target labels are consistent in left and right views (and
USC/CS/06-880 c ° 2006 arjf+ggm 3
across time, not visible in the figure).
1.3 System design
Usage of a relevant formalism greatly facilitates the design, study and implementation of complex dynamic
systems. This project adopted the Software Architecture for Immersipresence (SAI) framework [13].
SAI specifies a formal architectural style [12] whose underlying extensible data model and hybrid (shared
memoryandmessage-passing)distributedasynchronousparallelprocessingmodelfacilitatesthespecification
of asynchronous data stream processing software systems. A graph-based notation for architectural designs
allows intuitive system representation at the conceptual and logical levels, while at the same time mapping
closely to the physical level.
Figure 2 shows the conceptual specification graph for Stevi v.1 in the SAI notation. In the SAI model,
volatile data flows down streams (notation: thin lines) defined by connections between processing centers
called cells (notation: squares), in a message passing fashion. Repositories called sources (notation: circles)
holdpersistentdatatowhichconnectedcells(connectionnotation: thicklines)havesharedconcurrentaccess.
The directionality of stream connection between cells expresses dependency relationships. Dashed boxes in
the figure identify three main groups of components: (1) The Stereo input and pre-processing subsystem
regroups the multiple camera video input and subsequent pre-processing of the video frames flowing on the
stream; (2)theStereo multi-target detection and trackingsubsystemisexplainedindetailintheremainderof
thisdocument;(3)theVisualizationsubsystemprovidesavisualinsightintothetracker’sstatefordebugging
anddemonstrationpurposes. Thedesignisinherentlyparallel,andthemodularityoftheSAIstylefacilitates
design evolution.
Note that the graph presents an intuitive view of the system. Because the visual language in which it is
expressed has well defined semantics, the graph is also a formal specification (in this case at the conceptual
level).
The next four sections of this report explain in details the workings of the stereo multi-target tracking
subsystem: Section 2 introduces the principles of the feedback loop tracking pattern; Section 3 presents the
stereo mean shift-based tracking process; Section 4 describes the stereo target acquisition process; Section 5
addresses target prediction and tracker update mechanisms. Section 6 reviews system implementation and
performance. Finally, Section 7 presents a summary of contributions and directions for future work.
2 Integrated Detection and Tracking
Tracking is inherently a dynamic process. Making explicit the persistence of target trackers and the volatile
nature of target observations from data measurements naturally reveals the underlying feedback process.
2.1 A general tracking pattern
Thesubgraphdelimitedbytheboxlabeled Stereo multi-target detection and trackinginFigure2illustratesa
generaltrackingpatternthatcapturestheessenceofthefeedbackapproach. The Tracking sourceholdsalist
of target trackers. At each moment in time, the state of these trackers represents the latest known persistent
information for processing. A feedback loop anchored by the Predict targets cell and the Update trackers cell
accesses and maintains this information. The acquisition of each new data sample (stereo pair) triggers a
cycleoftheloop: Thepredictionprocess(Predict targets cell)generatesalistofexpectedobservationsbased
onthecurrenttrackers. Atrackingsubsystem(Track targetscell)usespredictionsandinputdatatoproduce
a list of confirmed tracked target observations. A residual analysis subsystem (Acquire targets cell) explains
input data not accounted for by the predictions, resulting in initial detection of additional targets. Finally,
an updating process (Update trackers cell) integrates confirmed target observations into the corresponding
trackers, and create new trackers for newly detected targets.
2.2 Properties of the pattern
This general pattern is inherently robust to some issues that are difficult to handle in static or sequential
approaches to online video tracking. Depending on specific design choices, a system based on this pattern
USC/CS/06-880 c ° 2006 arjf+ggm 4
Figure 2: Stevi v.1 conceptual level system architecture in SAI notation.
USC/CS/06-880 c ° 2006 arjf+ggm 5
may exhibit properties that are out of the reach of the individual algorithms involved for detection and
tracking. Because the Tracking and Acquisition subsystems operate on the same control loop, the tracking
process is naturally bootstrapped: unknown targets are automatically acquired. Furthermore, the absence
of observation for a given known target in a given input sample need not necessarily result in the loss of
the corresponding tracker (depending on the update policy). The system is thus resilient to a wide range of
disruptions and discontinuities in the input signal as well as errors in intermediate results caused by failures
of the individual algorithms.
2.3 Application in Stevi v.1
Stevi v.1 specifically addresses real-time multiple people detection and tracking in stereo input video.
Therefore the tracking and acquisition subsystems rely on algorithms specialized in face detection and skin
color tracking, with emphasis on the stereo integration, as described in the next sections. The algorithms
selected do not put any constraint on the motion of the targets or camera set.
The next three sections address respectively the tracking of known targets, new targets acquisition, and
target prediction and tracker update.
3 Target tracking
Givenasetofpredictedobservations(generatedfromthetrackers)andsomeinputdata,thetrackingprocess
produces a set of corresponding target observations that will be used to update the state of their respective
trackers. A variety of approaches can be applied to this task. In Stevi v.1, a mean shift-based approach
continuously adjusts the model to match the data. Stereo information is used to increase the robustness of
the system (cross-checking) and to dynamically adjust some parameters of the mean-shift.
3.1 Mean shift-based color tracking
The mean shift algorithm [6] is a robust non-parametric iterative gradient descent technique for finding the
mode of probability distributions. Convergence towards the mode of a probability distribution is achieved
by starting from a reasonable position, and iteratively shifting to a position computed as the weighted
mean of a kernel centered at the current position. This algorithm exhibits, under reasonable conditions,
very attractive convergence and robustness properties. In monocular mean shift-based color tracking [7], a
target is characterized by a color model (histogram), initialized from an external process, and its last known
position. Each input frame is converted into a probability distribution by attributing to each color pixel its
value in the target’s histogram. Starting from the last known position, mean shift is iterated to find the
mode of the distribution, estimated target location in the new frame. The search area can be restricted
around the last known position of the target, resulting in possibly large computational savings. The size of
the area to use for sampling the distribution is a fixed parameter. The Continuously Adaptive Mean SHIFT
(CAMSHIFT) algorithm [3] extends mean shift-based color tracking by continuously computing the size and
orientation of the tracked area, which is fed back to the mean shift process as the size of the sampling area
in the next frame. However, CAMSHIFT was not designed for, and does not perform well in situations
involving long range interaction and multiple targets. Instead, Stevi v.1’s tracking algorithm uses stereo
information computed for higher level structures (regions rather than pixels) to adjust the size of the mean
shift sampling window in successive frames.
3.2 Stereo mean shift-based tracking in Stevi v.1
TheinputtotheTrack targetcellconsistsofleftandrightinputcolordatainformationandasetofpredicted
target observations. The output of the mean shift tracking is a set of tracked target observations, i.e. target
observations updated according to the data using the predictions as starting point. In addition, the cell
produces left and right residual images, obtained from the original left and right input images, that are used
to reduce the search space for the target acquisition subsystem.
Each target prediction comprises two sets (left and right) of the following: color model (histogram),
current position and size. The tracking produces for each confirmed target an observation that comprises
USC/CS/06-880 c ° 2006 arjf+ggm 6
new positions and sizes in the left and right images. For each target, only a fixed number of iterations is
performed on each data sample. Typically, only one position update should be necessary, although a low
input sample rate relative to target image velocity might require more iterations. A key aspect is that mean
shift is never applied until convergence (i.e. iterated an undetermined number of times) for a given stero
image pair. The feedback loop in the tracking subsystem ensures that the new target positions are used as
starting point of their respective shifts in the next input samples. Note that if the input data is constant,
the tracked positions will converge to the same positions as they would in the traditional off-line iterative
approach.
Use of stereo redundancy confers added robustness to data noise and better adaptivity of the mean shift
process. It is important to emphasize that the tracking process does not involve any matching between left
and right information. A target observation explicitly encodes the conceptual relationship between its left
and right observation models (area and color histogram), which are enforced during tracking. Because the
cameras are close to each other and looking in parallel directions, the face area is constrained to have the
same size in both views. Furthermore, the shift observed in the right and left images should be similar
(although not strictly equal). Finally, the Euclidian distance between the centers of the tracked regions in
the left and right images estimates the disparity of the target. The size of the target area is computed as
a function of the disparity (from stereo calibration data), and thus indirectly fed back to the mean shift
process.
Figure 1 shows the tracked positions of three targets in both the left and right images. The apparent size
of the face area is estimated using stereo information. Color-based tracking is robust to head orientation (as
long as enough skin color remains visible).
4 Target acquisition
The target acquisition process analyzes input data not already explained by the tracking process. Stevi v.1
relies on a face detection module to acquire people’s heads. Although new target acquisition is a sparse
event, the need to continue searching for new targets in residual data (i.e. regions of the input images not
accounted for by the tracking module), as well as timely integration of new targets in the tracking process,
put efficiency constraints on this part of the system as well.
4.1 Face detection algorithm
The literature contains many approaches to face detection ([22, 14] offer recent surveys). In general,
appearance-based and learning-based techniques have produced the best results and show the most promise
to date [17]. Among these, the approach originally proposed by Viola and Jones [21] is the most compatible
with the real-time requirements of robotic vision. The recognition engine is a cascade of weak classifiers
utilizing Haar wavelet-like features. AdaBoost learning improves performance and efficiency of the learn-
ing process and of the resulting classifiers. This approach (with various subsequent extensions by other
authors [18, 19]) is suitable for real-time performance, and was thus adopted for face detection in Stevi v.1.
4.2 Stereo face detection in Stevi v.1
The Acquire targets cell combines the results of monocular face detection in left and right residual images
to infer the presence of new targets in the scene. The residual images, produced by the Track target cell,
invalidates images areas where faces are already being tracked. Cross validation in left and right views
increases the system’s robustness to false positives in the monocular face detection process. Simultaneous
detectionofoverlappingfaceareasinleftandrightimagesresultsintheinstantiationofatargetobservation
structure, with initial color histograms, positions and sizes in each view. The target area size is forced to
be identical in both views, consistently with a similar constraint enforced in the tracking. Performing target
detection for each input pair reduces the overall impact of detection failure in a particular input.
USC/CS/06-880 c ° 2006 arjf+ggm 7
Figure 3: ETRI robot platform: current prototype Weber-N (left) and consumer product design (right).
(Pictures courtesy of ETRI)
5 Target prediction and tracker update
The target prediction and tracker updating processes anchor the tracking feedback loop. Together with
appropriatetargettrackersstructures,theycouldimplementavarietyoftraditionalmethods,suchasKalman
filtering [2] and particle filtering [1]. Because Stevi v.1’s detection and tracking subsystem operates in 2D
imagespaceandundertheassumptionofasufficientinputframerate, simplicityandefficiencyarepreferred
to complex algorithms. The prediction process transfers the last know characteristics for each known target.
Similarly, the updating process is kept to the simplest minimum set of operations, by transferring the
observed characteristics of observed targets, creating new trackers for newly acquired targets, and removing
the trackers not supported by observation. This last policy means that the tracker for a target not observed
in a single input sample will be removed, and the target’s identity lost. A lost target will be re-acquired
as soon as the acquisition algorithm will allow, but will be assigned a new tracker (with a different label).
If one considers the output of the system to be the individual labeled trajectories for tracked targets (this
information is captured in the list of target tracker states at each moment), these trajectories would be
fragmented in the presence of occlusions or change in appearance. The implementation of more elaborate
prediction and update algorithms is one possible way to address some of these issues by allowing target
trackers to persist through accidental absence of observation, thus by making the tracking process locally
more robust.
6 Implementation and performance
Stevi v.1 was implemented in C++ using the Modular Flow Scheduling Middleware (MFSM) [10], an open
source architectural middleware that provides code support for SAI’s architectural abstractions. The graph
specification of Figure 2 constitutes a conceptual-level map of the code. Systems designed with SAI and
implemented with MFSM are automatically multi-threaded, and can directly take advantage of multiple
CPUs, hyper-threading and multi-core processor architectures. The MFSM project comprises a number of
opensourcefunctionalmodulesthatallowtofocuscodingeffortonprojectspecificcomponentsofthesystem.
Forexample,modulesforvideoinput,imagedisplay,andothergeneralpurposemodulesweredirectlyreused.
The basic computer vision algorithms underlying the system are established in the literature. Stevi v.1’s
implementation encapsulates face detection and color-based tracking functionalities provided by the Intel
OpenSourceComputerVisionLibrary(OpenCV)[4]. TheinitialSteviv.1systemwasdevelopedontheMS
Windows platform. The same system was ported to Linux with minimal effort, thanks to the cross-platform
MFSM code and module base, and the availability of Intel’s OpenCV for both platforms. Furthermore,
because of the modularity of the model, the development of subsequent evolutions of Stevi will directly
benefit from the work done on v.1.
Therunningsystemsreliablydetectandtrackpeopleatapproximately15frames-per-secondsoncontem-
porary hardware. Video input (images 320x240 pixels) is acquired by two color cameras (USB or Firewire)
mounted on a calibrated stereo rig. Stevi v.1 systems have run continuously for hours. A version using
a generic classifier cascade trained on frontal faces (as shipped with the MS Windows version of OpenCV
v.3.1) detects most new targets as soon as the faces become visible at the resolution required by the face
USC/CS/06-880 c ° 2006 arjf+ggm 8
detector. False positives, quite frequent with the monocular algorithm, are extremely rare occurences. Once
acquired, a face in motion can be tracked for several seconds to several minutes. Principal reasons for dis-
continuity are change of lighting/color (caused either by target motion in the environment or environmental
change), self occlusion. These situations expose the limited robustness of the largely skin-based color model,
and can be addressed by building more complex target models, or combining multiple detection algorithms,
which would allow more robust tracking. Occlusion by environment obstacles or other targets also causes
tracking discontinuity, due to the 2D nature of the detection data used in the tracking process. As already
mentioned, more elaborate prediction and update algorithms could make the tracking process locally more
robust. Another promising approach to improving overall system robustness is the addition of a higher level
3D tracking subsystem fed by the current detection and tracking system.
In all experiments, the most computationally expensive task is face detection, already performed on
reduced-resolutionimagesforefficiencypurposes. Thevisualizationsubsystem,usefulonlyfordemonstration
purposes, also consumes a non-negligible part of computation power. Stevi v.1 was integrated and tested
on the Weber-N personal service robotic platform developed by the Electronics and Telecommunications
Research Institute (ETRI, see Figure 3).
7 Conclusion
This report described the design and implementation of Stevi v.1, a computer vision system that performs
real-time people detection and tracking from stereo color video. Stevi v.1 illustrates the integration of
established monocular algorithms for face detection and color-based tracking in a general dynamic tracking
pattern whose core is a feedback loop. This pattern provides a general template for designing tracking
systems that exhibit resilience properties such as bootstrapping capability, robustness to input data noise,
adaptability to changing environment.
The complete stand-alone integrated system incorporates video capture, processing and result visual-
ization for evaluation purposes. It was designed using the SAI model. Parallel implementations on both
Windows and Linux platforms, using the open source MFSM architectural middleware, reliably detect and
track people at interactive rates.
SubsequentversionsofSteviwillachieveincreasedrobustnessthroughtheuseofmoreelaboratedynamic
targetmodels,andofmultipledetectionalgorithms. Anotherpromisingapproachtoimprovingoverallsystem
robustness is the addition of a higher level 3D tracking subsystem fed by the current detection and tracking
system.
7.1 Acknowledgments
This work was conducted in the context of a collaborative project with the Electronics and Telecommunica-
tions Research Institute (ETRI), a Korean non-profit, government-funded research organization.
References
[1] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp. A tutorial on particle filters for online
nonlinear/non-gaussian bayesian tracking. IEEE Transactions on Signal Processing, 2002.
[2] Y. Bar-Shalom and T. E. Fortmann. Tracking and data association. Academic Press, 1988.
[3] Gary R. Bradski. Computer video face tracking for use in a perceptual user interface. Intel Technology
Journal, Q2 1998.
[4] Gary R. Bradski. The OpenCV Library. Dr. Dobb’s Software Tools for the Professional Programmer,
November 2000.
[5] Cynthia Breazeal. Socially intelligent robots. ACM Interactions, 12(2):19–22, 2005.
[6] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 24(5):603–619, May 2002.
USC/CS/06-880 c ° 2006 arjf+ggm 9
[7] Dorin Comaniciu, Visvanathan Ramesh, and Peter Meer. Real-time tracking of non-rigid objects using
mean shift. In Proceedings of the Int. Conf. Computer Vision and Pattern Recognition, pages II: 142–
149, 2000.
[8] David J. Feil-Seifer and Maja J. Matari´ c. Socially assistive robotics. In Proceedings of the International
Conference on Rehabilitation Robotics, Chicago, Il, July 2005.
[9] T. Fong, I. Nourbakhsh, and K. Dautenhahn. A survey of socially interactive robots. Robotics and
Autonomous Systems, Special issue on Socially Interactive Robots, 42(3-4):143–166, 2003.
[10] Alexandre R.J. Fran¸ cois. Modular Flow Scheduling Middleware. mfsm.SourceForge.net.
[11] AlexandreR.J.Fran¸ cois.CAMSHIFTtrackerdesignexperimentswithIntelOpenCVandSAI.Technical
Report IRIS-04-423, Institute for Robotics and Intelligent Systems, University of Southern California,
July 2004.
[12] Alexandre R.J. Fran¸ cois. A hybrid architectural style for distributed parallel processing of generic
data streams. In Proceedings of the International Conference on Software Engineering, pages 367–376,
Edinburgh, Scotland, UK, May 2004.
[13] Alexandre R.J. Fran¸ cois. Software Architecture for Computer Vision. In G. Medioni and S.B. Kang,
editors, Emerging Topics in Computer Vision, pages 585–654. Prentice Hall, Upper Saddle River, NJ,
2004.
[14] E. Hjelmas and B.K. Low. Face detection: A survey. Computer Vision and Image Understanding,
83(3):236–274, September 2001.
[15] Tony S. Jebara and Alex Pentland. Parameterized structure from motion for 3d adaptive feedback
tracking of faces. In Proceedings of the International Conference on Computer Vision and Pattern
Recognition, pages 144–150, 1997.
[16] Kazuhiko Kawamura, Robert T. Pack, and Moenes Iskarous. Design philosophy for service robots.
In Proceedings of theIEEE International Conference on Systems, Man and Cybernetics: ’Intelligent
Systems for the 21st Century’., volume 4, pages 3736–3741, Vancouver, BC, Canada, October 1995.
[17] Stan Z. Li and Juwei Lu. Face detection, alignment, and recognition. In G. Medioni and S.B. Kang,
editors, Emerging Topics in Computer Vision, pages 385–455. Prentice Hall, Upper Saddle River, NJ,
2004.
[18] S.Z. Li, L. Zhu, Z. Zhang, A. Blake, H. Zhang, and H. Shum. Statistical learning of multi-view face
detection. In Proceedings of the Europeean Conference on Computer Vision, page IV:67 ff., 2002.
[19] R. Lienhart, A. Kuranov, and V. Pisarevsky. Empirical analysis of detection cascades of boosted
classifiers for rapid object detection. In Bernd Michaelis and Gerald Krell, editors, Proceedings of the
25thDAGMSymposiumonPatternRecognition,pages297–304,Magdeburg,Germany,September2003.
Springer Verlag.
[20] FrancescMoreno,AdriaTarrida,JuanAndrade-Cetto,andAlbertoSanfelin. 3dreal-timeheadtracking
fusing color histograms and stereovision. In Proceedings of the International Conference on Pattern
Recognition, pages I: 368–371, 2002.
[21] Paul Viola and Michael J. Jones. Robust real-time face detection. International Journal of Computer
Vision, 57(2):137–154, 2004.
[22] M.H. Yang, D.J. Kriegman, and N. Ahuja. Detecting faces in images: A survey. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 24(1):34–58, January 2002.

Linked assets

Computer Science Technical Report Archive

Conceptually similar

PDF

USC Computer Science Technical Reports, no. 889 (2007)

PDF

USC Computer Science Technical Reports, no. 593 (1994)

PDF

USC Computer Science Technical Reports, no. 794 (2003)

PDF

USC Computer Science Technical Reports, no. 914 (2010)

PDF

USC Computer Science Technical Reports, no. 867 (2005)

PDF

USC Computer Science Technical Reports, no. 578 (1994)

PDF

USC Computer Science Technical Reports, no. 549 (1993)

PDF

USC Computer Science Technical Reports, no. 575 (1994)

PDF

USC Computer Science Technical Reports, no. 572 (1994)

PDF

USC Computer Science Technical Reports, no. 574 (1994)

PDF

USC Computer Science Technical Reports, no. 773 (2002)

PDF

USC Computer Science Technical Reports, no. 669 (1998)

PDF

USC Computer Science Technical Reports, no. 639 (1996)

PDF

USC Computer Science Technical Reports, no. 787 (2003)

PDF

USC Computer Science Technical Reports, no. 874 (2006)

PDF

USC Computer Science Technical Reports, no. 660 (1997)

PDF

USC Computer Science Technical Reports, no. 804 (2003)

PDF

USC Computer Science Technical Reports, no. 970 (2017)

PDF

USC Computer Science Technical Reports, no. 931 (2012)

PDF

USC Computer Science Technical Reports, no. 595 (1994)

Description

Alexandre R.J. Francois, Gerard G. Medioni. "A vision system for personal service robots: Resilient detection and tracking of people." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 880 (2006).

Asset Metadata

Creator Francois, Alexandre R.J. (author), Medioni, Gerard G. (author)

Core Title USC Computer Science Technical Reports, no. 880 (2006)

Alternative Title A vision system for personal service robots: Resilient detection and tracking of people (title)

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)

Tag OAI-PMH Harvest

Format 9 pages (extent), technical reports (aat)

Language English

Unique identifier UC16270908

Identifier 06-880 A Vision System for Personal Service Robots Resilient Detection and Tracking of People (filename)

Legacy Identifier usc-cstr-06-880

Format 9 pages (extent),technical reports (aat)

Rights Department of Computer Science (University of Southern California) and the author(s).

Internet Media Type application/pdf

Source 20180426-rozan-cstechreports-shoaf (batch), Computer Science Technical Report Archive (collection), University of Southern California. Department of Computer Science. Technical Reports (series)

Access Conditions The author(s) retain rights to their work according to U.S. copyright law. Electronic access is being provided by the USC Libraries, 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.

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Repository Email csdept@usc.edu

Inherited Values

Title Computer Science Technical Report Archive

Description Archive of computer science technical reports published by the USC Department of Computer Science from 1991 - 2017.

Coverage Temporal 1991/2017

Repository Email csdept@usc.edu

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)