Computer Science Technical Report Archive

USC Computer Science Technical Reports, no. 739 (2001)

(USC DC Other)

USC Computer Science Technical Reports, no. 739 (2001)

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Alternativ e T ec hniques for the Ecien t A cquisition of Haptic Data
Cyrus Shahabi, Greg Barish, Mohammad R. K olahdouzan, Didi Y ao,
Roger Zimmermann, Kun F u, Lingling Zhang
In tegrated Media Systems Cen ter and Departmen t of Computer Science
Univ ersit y of Southern California
Los Angeles, CA, 90089
[shahabi, gbarish, kolahdoz, didiyao, rzimmerm, kfu, linglinz]@usc.edu
No v em ber 15, 2000
Abstract
Immersiv e en vironmen ts are those whic h aim to surround users in an articial w orld. These en vironmen ts
consist of a comp osition of v arious t yp es of immersidata: unique data t yp es that are com bined to render a virtual
exp erience. A cquisition, for storage and future querying, of information describing sessions in these en vironmen ts
is c hallenging b ecause of the real time demands and sizeable amoun ts of data to b e managed. In this pap er, w e
compare tec hniques for ac hieving the ecien t acquisition of one t yp e of immersidata, the haptic data t yp e, whic h
describ es the mo v emen t, rotation, and force asso ciated with user-directed ob jects in an immersiveen vironmen t. In
addition to describing a general pro cess for real-time sampling and recording of this t yp e of data, w e prop ose three
distinct sampling strategies: xe d, gr oup e d and adaptive. W e conduct sev eral exp erimen ts with a real haptic device,
Cyb erGr asp,to ev aluate the eciency , accuracy and complexit y of our prop osed tec hniques. Our results showthat
xed sampling, acquiring data at the maxim um sampling rate p ossible, is easy to implementbut requires high
This researc h has b een funded in part bythe In tegrated Media Systems Cen ter, a National Science F oundation Engineering Researc h
Cen ter, co op erativ e agreementno. EEC-9529152, b y NASA/JPL con tract no. 961518, D ARP A and USAF under agreementno. F30602-
99-1-0524, and unrestricted cash/equipmen t gifts from In tel, IBM, NCR and SUN.
1
amoun ts of storage and bandwidth. Group ed sampling, whic h iden ties and records at the optimal rates for eac h
sensor group on haptic devices, is more ecien t but can lead to o v er-sampling or under-sampling in certain scenarios.
Finally,w e found that adaptiv e sampling, p erio dically optimizing the sampling rate at run-time for the individual
sensors on a giv en haptic device, requires more run-time complexit y but pro vides substan tial sa vings in b oth storage
and bandwidth, without signican t loss of accuracy . As immersiveen vironmen ts b ecome more complex and con tain
more haptic sensors, tec hniques suc h as adaptiv e sampling can b e useful to w ard scaling real-time acquisition of data
in these en vironmen ts.
1 In tro duction
Immersiveen vironmen ts are those whic h aim to surround users in a com bination of real and articial w orlds, connecting
a p erson with other p eople, ob jects, places and databases through an augmen ted or virtual realit y exp erience. These
en vironmen ts are comp elling b ecause they enable applications that sim ulate realit y . F or certain application domains,
suc h as virtual training, an immersiv e en vironmen t dramatically impro v es the practicalit y and cost-eectiv eness of
those applications. Learning to p erform surgery in a virtual en vironmen t, for example, allo ws aspiring surgeons to
exibly and ecien tly practice their tec hniques - without requiring an actual b o dy during the pro cedure.
Immersiv e en vironmen ts consist of a comp osition of v arious t yp es of immersidata [SBE+99 ], unique data t yp es
that are com bined to render a virtual exp erience. One suc h immersidata t yp e is the haptic data t yp e, information that
describ es the mo v emen t, rotation, and force asso ciated with user-directed ob jects in an immersiveen vironmen t. Haptic
data is pro duced byin terfaces that connect the user to the immersiveen vironmen t. These in terfaces are conceptually
similar to mice in that they translate user motion to avisualen vironmen t. Ho w ev er, they are m uc h more complex
than mice b ecause they are often equipp ed with man y sensors, th us rotations, forces, and three-dimensional p ositioning
asso ciated with h uman activit y can b e captured. They also supp ort force-feedbac k, so that users w earing the device
can exp erience the sensation of in ten tionally touc hing ob jects in that virtual en vironmen t, as w ell as unin ten tional
collisions.
Curren tly,w e are studying the problem of ecien t, real-time haptic data acquisition. A cquisition is critical b ecause
it allo ws the data to b e stored and th us enables sev eral imp ortan t mo des of op eration on that stored data. One suc h
mo de is playb ack, whichallo ws a previously recorded session to b e repla y ed (re-exp erienced). Th us, virtual training
applications (suc h as those for virtual surgery or pilot training) can b e built. Another mo de of op eration is that of
simple querying of the data. Numerous rep orting-st yle applications can b e built from this, suc h as those that describ e
ho w man y times a user touc hed v arious ob jects in the scene or those that graphically render a diary of the user’s
exp erience. A third mo de of op eration in v olv es detailed analysis of the stored data. This is more akin to the goals
2
of data mining, where certain trends or rules of the data can be learned for use in adv anced applications suc h as
customization or p ersonalization. F or example, if a system can learn that pilots b eing trained w ere sp ending a lot
of time adjusting a particular con trol on the virtual aircraft, it migh t indicate that the virtual con trol is dicult to
use. Corresp ondingly , the system designers can tak e a more detailed lo ok at the design of that con trol and ensure the
accuracy of its 3-D mo del and rendering.
While haptic data acquisition is useful, it is not easy . It is particularly c hallenging b ecause of the real-time demands
of sim ultaneously recording m ultiple sensors, main taining the accuracy of recording, and handling the v oluminous
data pro duced during a session of reasonable length. Researc h related to other t yp es of iso c hronous, con tin uous media
(suc h as video [KK97 ],[LHHC96 ]) ha veencoun tered similar c hallenges. Ho w ev er, these t yp es of media do not ha vethe
inherentin ternal structure or spatio-temp oral prop erties of haptic data. These prop erties directly aect the acquisition
metho dologies w e describ e.
One approachw e are exploring deals with reducing the sampling rate for a giv en sensor (or group of sensors) to
alev el that reduces the amoun t of data w e need to collect without sacricing accuracy . The general in tuition is: for
certain sessions or parts of sessions, man y sensor v alues ma y not c hange or ma y not exhibit a sp ectrum of v alues that
demands a high sampling rate. F or these cases, it w ould seem useful to reduce the sampling rate to a m uchlo w er lev el.
The problem is howto iden tify that lev el. In this pap er, w e describ e run-time and oine approac hes that iden tify
sampling rates based on the sensor (or sensor group) and the seman tics of the session. Our tec hniques rely on the
Nyquist theorem to iden tify sampling rates that do not sacrice accuracy . Through sev eral exp erimen ts with a real
haptic device, Cyb erGr asp, w e compare our approac hes with the default sampling rate and discuss tradeos b et w een the
v arious approac hes and corresp onding implications. Our most ecien ttec hnique, adaptiv e sampling, reduces storage
and bandwidth requiremen ts b y 76% to 97% while reconstructing the original data with 85% to 99% accuracy . W e
also sho w that the sa vings in storage is b et w een 2 to 20 times b etter than applying traditional compression algorithms
to the data.
Sp ecically , this pap er mak es the follo wing con tributions:
Describ es a general pro cess for sampling and recording haptic data in a real-time en vironmen t.
Sho ws ho w the Nyquist theorem can b e used in the acquisition pro cess to iden tify the optim um sampling rate
for haptic devices.
Prop oses t wono v el acquisition tec hniques more applicable to immersiveen vironmen ts.
Compares these tec hniques in terms of their trade-os b et w een eciency , computational needs and accuracy .
Demonstrates that our session and sensor dep enden t sampling tec hnique reduces space requiremen t signican tly
as compared to con v en tional compression algorithms.
3
Finds that applying extra, traditional data compression do es not alw a ys further reduce the size of the acquired
data.
The remainder of the pap er is organized as follo ws: in Section 2, w e describ e haptic data and the device weused in
our exp erimen ts in more detail. Section 3 in tro duces the three sampling tec hniques and discusses their relativ e merits.
W e address the related w ork in Section 4. The exp erimen tal results and discussion are pro vided in Sections 5 and 6
resp ectiv ely . Finally,w e iden tify our conclusions and future w ork in Section 7.
2 Haptic Data Denition
Haptic data consists of a series of sensory v alues measured at some p ointintime. Some of these sensory v alues maybe
pro duced b y sensors that b elong to the same logical group. F or example, a glo vein terface mightha v e sensors for eac h
nger. When a user touc hes or grabs an ob ject, this logical group migh t be asso ciated with related sensory v alues.
T oda y , there are commonly three basic t yp es of sensors asso ciated with haptic devices: those that measure p osition,
those that measure rotation, and those that measure force lev els. Ho w ev er, there are man y other t yp es of sensors that
exist. F or example, a haptic glo v e or exosk eletal device that slips on to the hand migh t measure nger angles. Also,
the n um b er and placemen t of sensors can aect the accuracy of represen ting haptic sessions. A cquiring haptic data
in v olv es sampling it at v arious p oin ts in time. In general, it w ould seem that the higher the sampling rate, the more
precise a session can b e captured. Ho w ev er, dep ending on the session, there are often p oin ts of diminishing returns on
suc h rates. Finding the optimal sampling rate is the general fo cus of Section 3.
2.1 Cyb erGrasp
Haptic device dev elopmen t is still in the v ery early stages. Not man y pro ducts are a v ailable; those that are can b e quite
exp ensiv e. Weha v e fo cused our study of haptic data acquisition on one set of the a v ailable devices: the Cyb erGrasp
exosk eletal in terface and accompan ying Cyb erGlo v e (b oth sho wn in Figure 1a), from Virtual T ec hnologies, along with
aP olhem us F astrak p osition trac king device. The detailed description of 33 sensors included with these devices are
depicted in T able 1.
When users attac h the p osition trac k er and w ear the Cyb erGrasp and Cyb erGlo v e on one hand, they are able
to mo v e a virtual hand in an immersiveen vironmen t. Th us, users can touc h or grab other ob jects and exp erience
the ph ysical attributes of those ob jects. They can also exp erience resistance forces, suc h as those asso ciated with
holding and releasing an ob ject. F or example, they can exp erience pic king up and releasing a virtual ball or the
resistance asso ciated with pushing against a w all. Figure 1b sho ws a user w earing t w o Cyb erGlo v e/Cyb erGrasp
4
Sensor Num ber Sensor Description Sensor Num ber Sensor Description
1 th umbrollsensor 14 ring outer join t
2 th um b inner join t 15 ring-middle ab duction
3 th um b outer join t 16 pinky inner join t
4 th um b-index ab duction 17 pinky middle join t
5 index inner join t 18 pinky outer join t
6 index middle join t 19 pinky-ring ab duction
7 index outer join t 20 palm arc h
8 middle inner join t 21 wrist exion
9 middle middle join t 22 wrist ab duction
10 middle outer join t 23,24,25 x,y ,z lo cation
11 middle-index ab duction 26,27,28 x,y ,z ab duction
12 ring inner join t 29 to 33 forces for eac h nger
13 ring middle join t
T able 1: Cyb erGrasp sensors
a. The Cyb erGrasp and Cyb erGlo v e b. T ranslation of user action in to virtual en vironmen t
Figure 1: Cyb erGrasp (gure reprin ted from Virtual T ec hnologies)
devices, p erforming an action in the real w orld, and w atc hing its translation to the on-screen en vironmen t. The
complete conguration allo ws a full-range of motion with force feedbac k to prev en t ngers from p enetrating or crushing
a virtual ob ject.
2.2 Characteristics of Cyb erGrasp
Cyb erGrasp in terfaces t w o sub classes of haptic data: grasping data and kinesthetic data. Grasping data is a set of 28
oating p oin tv alues corresp onding to 22 angles for the 22 degrees of freedom of the hand, 3 v alues for hand co ordinates
in 3D Cartesian space (x, y, z ) with a range of [-120", 120" ], and 3 angles for hand orien tation (Euler-based rotational
co ecien ts) with a range of [- , ]. These v alues can b e captured b y monitoring the amoun t of stretc hing, b ending,
and Euclidean distances of the b ers within the Cyb erGrasp ngers as w ell as the trac king device lo cated on the cen ter
5
a. Grasping Data b. Kinesthetic Data
Figure 2: Sub classes of the haptic data t yp e for Cyb erGrasp
of the palm. As sho wn in Figure 2a, the 22 angles for the 22 degrees of freedom with a range of [- , ] consist of:
14 angles, i
(1 i 14 ), corresp onding to the 14 join ts of the v e ngers of a hand,
4 angles, j
(1 j 4 ), b et w een the v e ngers,
1 angle, 1
, for the second degree of freedom of the th um b
1 angle, 2
, for the one degree of freedom of the cen ter of palm, and
2 angles, (not sho wn in the Figure 2) for wrist exion and ab duction
Finally , the kinesthetic data is a set of 5 v alues with a range of [0, 1], represen ting the force applied to eac h nger,
captured through monitoring Cyb erGrasp tendons. Eachv alue is in the range of [0,1], whic h is the equiv alentof F
x
,
F
y
,F
z
(see Figure 2b).
3 Data A cquisition
Amajor c hallenge in haptic data acquisition is deciding ho w often to acquire the sensory v alues. A naiv e approac h
mayin v olv e sampling the sensor status as fast as the acquisition system (b oth hardw are and soft w are) can op erate.
The in tuition is that the more samples w e collect, the higher the accuracy . On the other hand, due to either device
limitations or the nature of h uman motion, the v alue/status of a sensor mightnot c hange as fast as the system samples.
Hence, the sampling rate will b e higher than necessary , resulting in w asted storage space and bandwidth. Iden tifying a
lo w er sampling rate will enable haptic data acquisition to scale-up to man y concurren t users and sensors, th us enabling
in tegration in en vironmen ts with lo w bandwidth net w orks. A cquiring haptic data b ecomes ev en more complicated if
6
w e consider that the optimal sampling rate also dep ends on b oth the sensor b eing sampled and the immersiv e session
b eing recorded. W e conducted some studies to understand dieren t factors that impact the sampling rate. In this
section, wereportonthe results of our studies. W e plan to extend this study to in v estigate an adaptiv e tec hnique
that iden ties the optimal sampling rate on-the-y based on b oth previous kno wledge of the session and sensors and
analysis of a windo w of buered data. The results rep orted here are our attempt to demonstrate the b enets of suc h
an adaptiv e tec hnique.
In our exp erimen ts, w e used the Cyb erGrasp Soft w are Dev elopmen t Kit (SDK) to write handlers that record sensor
data whenev er a particular sampling in terrupt w as called. The rate at whic h these handlers w ere called -th us, the
maxim um rate w e could sample - v aried as a function of the CPU sp eed.
T o sample and record data async hronously,wedev elop ed a simple m ulti-threaded double buering approac h. One
thread w as asso ciated with answ ering the handler call and cop ying sensor data in to a region of system memory . A
second thread w ork ed async hronously to pro cess and store that data to disk. The CPU w as nev er 100% busy during
this pro cess, so w e do not b eliev e our recording strategy in terfered with the rendering pro cess itself. F urthermore,
there is ob vious ro om for optimization here: w e could run our exp erimen ts on dual-pro cessor mac hines and wecould
also adjust the thread priorities for the second thread.
3.1 Basic Sampling Metho dology
Our sampling metho dology describ ed b elo w assumes familiarit y with Nyquist-based sampling tec hniques
1
. Based on
the Nyquist theory,w e need to sample a signal with a rate t wice as fast as the maxim um frequency in the signal in
order to be able to fully reconstruct it. This means that the required sampling rate should be t wice the maxim um
frequency in the sp ectrum of the signal (the frequency domain represen tation of the original signal):
r
nyquist
=2f
max
In order to lo cate f
max
,w e need to apply a discrete F ourier transform on the input signal consisting of x
0
... x
N
,
where N is the n um b er of samples tak en for a giv en windo w (a series of windo ws comp ose a session). Sp ecically:
X
k
=
N
X
i=1
x
i
:e
j 2 (k 1)(
i 1
N
)
1 k N; j =
p
1
where the absolute v alues of X
k
sp ecify the energy of the original signal. This is v alid if the input signal is deterministic.
1
The Nyquist Theorem is a fundamen tal signal pro cessing tec hnique describ ed b y [Nyq24] and [SW48].
7
But for cases where the input signal is random, w e need to apply the F ourier transform on the auto-correlation of the
input signal. The auto-correlation of a random signal pro vides a more precise measure on ho w fast the signal c hanges.
The auto-correlation function can b e dened based on the con v olution of the signal with its time-rev ersal as:
Auto
C or r el ation(x)= x(n) x( n)
where the con v olution of t w o signals is dened as:
x(n) y (n)=
N 1
X
m=0
x(m):y (n m)
whic h pro duces 2N-1 symmetric v alues for an input signal with N v alues. It is practical to in tro duce a cut-o frequency ,
f
cut of f
, where the signal has a negligible amoun t of energy (e.g., 1%) for frequencies greater than that, and use the
f
cut of f
instead of f
max
. This in tro duces the notion of c ondenc ethr eshold (CT), whic h is the p ercen tage of the energy
of the signal that is signican t. This threshold is application sp ecic and usually has v alues b et w een 70% and 99%.
W e study the impact of the condence threshold in Section 5. Once the discrete F ourier transform of the signal (or
its auto-correlation) is computed, the total energy of the signal can b e found as:
T otal
E ner g y =
N
X
k =1
X
k
Using the total energy of the signal and selecting the desired condence threshold for our application, w e can
compute f
cut of f
as:
f
cut of f
= M wher e :
(
(C onf idence
T hr eshold T otal
Energy ) P
M
k =1
X
k
(C onf idence
T hr eshold T otal
Energy ) P
M +1
k =1
X
k
This metho d is applicable to b oth con tin uous and discrete signals. F or a con tin uous signal, f
cut of f
is used as
the sampling rate. A discrete signal with rate r can b e do wn-sampled with the ratio of r/f
cut of f
, whic h pro duces
a r e duc e d r ate signal with rate f
cut of f
. In practice, the input signals from mec hanical instrumen ts ha v e inseparable
white noise. T o reduce the eect of the white noise, the input signal is passed through a lo w-pass lter b efore sampling
(or do wn-sampling). The reduced rate signal should b e up-sampled with the ratio f
cut of f
/r to pro duce a r e c onstructe d
signal with rate r. Wecan ev aluate the accuracy of the f
cut of f
and condence threshold v alues b y using the results of
the minimum squar e estimate function as applied to the original and reconstructed signals. Sp ecically , the denition
8
of the function is:
M inimum S q uar e E stimate(x; y)=
P
N
i=1
(x
i
y
i
)
2
P
N
i=1
(x
i
)
2
where x
0
... x
N
and y
0
... y
N
are the v alues of the original and reconstructed signals. The optimal v alue for the
minim um square estimate function is 0, when x
i
=y
i
, a condition that the Nyquist theorem guaran tees will b e met
when f
max
is used to do wn-sample the original signal. Instead, sampling with rate f
cut of f
in tro duces v alues greater
than 0. Dep ending on the application, a v alue b et w een 0.01 to 0.2 for the minim um square estimate function determines
an appropriate c hoice for the condence threshold (and hence f
cut of f
).
3.2 Fixed Sampling (FS)
W e dene xe d sampling as the case that sampling is done for all haptic device sensors at a constan t rate. There
are t w o approac hes that can b e used for xed sampling. One is to use the maxim um sampling rate, r
max
,allo w able
b y the SDK (as a function of the CPU sp eed). While this is easy to program, it is also the most w asteful of the
sampling approac hes describ ed since it records data for eac h sensor at ev ery opp ortunit y , regardless of sensor t yp e or
the seman tics of the session. In general, this t yp e of sampling is xed for all haptic devices connected to the same
mac hine: it simply uses the maxim um sampling rate allo w able b y the SDK and mac hine CPU. In our setup r
max
=
80 Hz.
A more ecien t form of xed sampling - denoted as mo die dxe d sampling (MFS)- in v olv es nding the minim um
sampling rate r
0
required b y all of the haptic device sensors, and then use that v alue as the sampling rate. As describ ed
ab o v e, w e use our basic metho dology to iden tify r
0
for the en tire sensor set on a haptic device, and then use that as
the sampling rate. Our exp erimen ts in Section 5 determine r
0
as 67 Hz. There is one disadv an tage to this approac h:
weneedto iden tify r
0
b efore w e can start sampling at that rate. This means that wem ust ha v e tr aining sessions for
the haptic device w e are in terested in recording. Ho w ev er, once w e iden tify r
0
, w e can p oten tially sa v e substan tial
amoun ts of disk space and bandwidth since the amoun t of data recorded p er second can b e signican tly less than that
allo w able b y r
max
. Note that using r
0
instead of r
max
in our setup pro vides only 16% reduction for bandwidth and
storage requiremen ts, but setups with faster CPUs (hence higher v alues for r
max
) result in b etter impro v emen ts.
3.3 Group ed Sampling (GS)
Wein tro duce the gr oup e d sampling approac h, whic h tak es the mo died xed sampling approac h one step further and
iden ties the minim um sampling rate for dieren t groups of sensors on an immersidata device. The in tuition here is
that devices suc h as Cyb erGrasp ha v e sev eral sensors, and in man y cases these sensors can be mapp ed to distinct
9
groups
2
. Th us, w e can isolate sampling rates for eac h group and th us acquire data at dieren t rates, based on group
mem b ership.
F or example, on the Cyb erGrasp there are v eforce actuators, one for eachnger. In man y sessions, users are
mo ving around in an immersiveen vironmen t, but not holding or colliding with ob jects. In these cases, the force data
from eachof these sensors is lik ely to be the same (assuming noise can be ltered out) and th us this group can be
sampled far less frequen tly . The adv an tage here is an impro v emento v er the basic mo died xed sampling b y reducing
storage and transmission requiremen ts further while main taining accuracy . Ho w ev er it still retains the disadv an tage
of requiring training sessions in order to iden tify sampling rates for eac h of the groups.
The dicult y in pursuing a group ed sampling strategy in the general case is iden tifying the groups. Using h uman
in telligence, w e can easily distinguish b et w een nger force sensors and trac k er sensors and understand wh y they are
dieren t. Ho w ev er, for an arbitrary set of immersidata sensors, it ma y not b e easy to iden tify these groups. In addition,
our o wn in tuition ab out natural groups in some cases ma y not b e correct or meaningful. P erhaps the seman tics of the
sessions w e record indicate that there is no need to distinguish b et w een some natural groups, and to instead collapse
them in to a single group. With these issues in mind, w e consider automating group ed sampling as an approachto
perform standard cluster analysis tec hniques o v er the set of optimal rates, obtained during the training session for
eac h sensor, and form groups based on those results.
3.4 A daptiv e Sampling (AS)
Finally , w e dene adaptive sampling as a dynamic form of sampling that is applied to eac h sensor. With adaptiv e
sampling, w e are prop osing an approac h that w ould iden tify an optim um rate r
ij
for eac h sensor i during a giv en
windo w j of the session. Th us, sampling rates for eac h of the sensors are dynamic and corresp ond to the seman tics of
the session.
Ob viously , the main adv an tage here is the optimalit y asso ciated with this approac h. A daptiv e sampling could
poten tially reduce bandwidth and storage requiremen ts to far lo w er lev els than what xed or group ed sampling did,
enabling a m uc h wider range of sensors to participate in co op erativ e immersipresence. F urthermore, unlik e group ed
sampling, a training session do es not dictate the lifelong rate of a particular set of sensors. Rather, the rates c hange
as the nature of the sessions c hange. This mak es the approachmore durable than group ed or ev en mo died xed
sampling, where training sessions ma y cause o v er-sampling or under-sampling due to the p ossible dierences b et w een
real and training sessions.
Ho w ev er, adaptiv e sampling also demands a more complex implemen tation approac h. Using a double-buering
2
F or the purp ose of this pap er, w e use the terms sensor typ es and sensor gr oups in terc hangeably .
10
approac h, w e prop ose ha ving a r e c or ding thr e ad that samples at the maxim um rate p ossible and cop y the information
to memory . Then a stor age thr e ad p erforms the basic sampling metho dology to iden tify the Nyquist sampling rates,
r
ij
, for eac h of the sensors at p erio dic times, do wn-sample and store/transmit p er those rates. This implies t w o minor
requiremen ts ab out the immersidata acquisition system: (a) the CPU m ust b e sucien tly a v ailable to p erform these
calculations and (b) there m ust b e enough memory on the system to handle the buering and calculation requiremen ts
asso ciated with the set of immersidata sensors. Pro cessing input data with n v alues requires enough memory for 5n
v alues: n for the original data, 2n for auto-correlation of the original data and 2n for the F ourier transform of the
auto-correlation. The memory required b y the auto-correlation or F ourier transform v alues can b e recycled for the n
v alues for the ltered and reduced size data.
While b oth of those are not ma jor issues, a cause for concern is the dela y (sp ecically the length of the sampling
windo w) in tro duced b y this pro cess. T o do ecien t haptic data acquisition, it is useful to buer a fairly large amoun t
of data b efore p erforming calculations on that data to w ards iden tifying Nyquist rates mainly b ecause of the random
nature of the data. F or example, with our Cyb erGrasp system, weha v e found that sampling at 80 Hz implies that w e
need to collect data for a p erio d of around 10-15 seconds b efore pro cessing, transmitting and storing the data. Th us,
the buering pro cess means that some degree of the real-time acquisition is sacriced.
4 Related W ork
Haptic data generally has been classied as tactile data, grasping data, or force feedbac k data. T actile sensing
in v olv es the dynamics of the h uman ngertips to b etter understand the in terface b et w een skin and ob ject. Previous
studies observ e the compressibilityand shape c hanges of the ngertip and determined the highest c hange in v olume
of the ngertip to be 5% [SGD92]. Fingertip v olume w as also measured when surfaces with inden tions and shear
displacemen ts w ere applied against the ngerpad. Besides measuring the deformation of the ngertip when applied
to shap e inden tors, the gro wth and motion of con tact regions are measured along with the asso ciated force v ariations
o v er time bet w een the h uman ngerpad and the transparen t test ob jects whose micro-texture, shap e or softness is
v aried in a con trolled manner. Circular and rectangular inden tors w ere used to deform the ngerpad [RDS94], [RS95 ].
The deformation w as trac k ed b y measuring the displacemen ts of mark ed areas on the ngerpad. This exp erimen tal
data w as compared to data obtained through sim ulated ngerpad deformation and the comparisons w ere fa v orable.
Studying ngerpad deformation can help determine what areas of the skin are making con tact with ob jects and with
what force is exerted on to that ob ject. Certain areas of the skin will exert greater force th us making the feeling
dieren tthanwith ligh ter force. Tin y arra ys of cilia on an electronic c hip device can mimic the texture and force of an
11
ob ject making con tact with the skin, but stim ulating the correct cilia to induce a feeling is a c hallenge. The enormous
amoun ts of cilia stim ulus data, whichco v er a wider area of skin, is another c hallenge.
An example of a haptic in terface is the sensing c hair b eing dev elop ed at MIT Media Lab and Purdue Univ ersit y .
This c hair is moun ted with con tact sensors whic h monitors the p osture and p osition of someone sitting in the c hair
[T an00 ]. A p erson’s p osture is a go o d indicator of ho w tired or stressed that p erson is feeling. Also, analyzing the
b one structure of the bac k and p elvis area could b e a go o d iden tication mec hanism. Dep ending on the n um ber of
sensors in the c hair, the actual data acquisition ma y b e simpler than in terpreting the data.
Muc h of the prior w ork has in v olv ed in v estigating the PHANT oM force-feedbac k device from SensAble T ec hnologies,
Inc. The PHANT oM device allo ws a user to feel a virtual ob ject using a rigid stic k. The relativelo w-cost and the
a v ailabilit y of applications mak es the PHANT oM a suitable testb ed for researc h.
Haptic w ork at the MIT Media Labs fo cuses on rendering static and dynamic holographic images and videos.
With pre-stored 3D holograph y data, they are able to render a virtual hologram where a force-feedbac k PHANT oM
device is used to touc h the hologram. Also, the hologram can be rendered as a dynamic video image where a
similar PHANT oM device can be used to carv e and mo dify the image close to real-time. The mo died result can
later b e prin ted using a 3D prin ter [PP99]. An in teresting problem encoun tered b y the holograph y haptic in terface
is that mismatc hes b et w een vision and feeling can b e v ery noticeable. When touc hing the visual hologram with the
PHANT oM device, if the PHANT oM device do es not exhibit an y force-feedbac k precisely when the ey e sees that the
device mak es con tact then this is an ob vious miscue.
Man y of the p oten tial applications of haptic in terfaces in v olv e training and sim ulation. The Virtual RealityDen tal
T raining System Den tal sim ulator w as dev elop ed using a PHANT oM with four tips whic h sim ulates actual den tal
instrumen ts [AR99]. This is used to train studen ts to explore materials similar to to oth enamel. The PHANT oM has
also been used to sim ulate a to ol used for arthroscopic surgery [Mor98 ]. This giv es the studen ts the exp erience of
encoun tering ligamen ts and brous materials when p erforming surgery . Haptics can also pla y a role in graphics and
arts. T raining studen ts to p erform certain brush strok es or molding a sculpture a certain w a y can all b e demonstrated
using haptics in terfaces [Sjo97 ], [SA V98]. Computer-aided design and animation manipulation are areas where haptics
can enhance and facilitate curren t pro cesses. Here at IMSC at USC, researc h is b eing conducted to dev elop a haptic
m useum [MSH+00 ] where m useum visitors can explore m useum pieces b y touchas w ell as sigh t. This pro vides m useum
visitors with more information ab out a particular piece without inducing w ear and tear to the original piece.
12
Figure 3: Flowc hart for pro cessing Cyb erGrasp data
No. Description
1 Still hand with no mo v emen t
2 Th umbmo v es up and do wn once with hand in a closed-st
3 P oin ter and middle ngers mo v e up and do wn once with hand in a closed-st
4 Hand op ens and closes once with all ngers
5 Same as Session 2 with up and do wn motion rep eated 5 times
6 Same as Session 2 with up and do wn motion rep eated 5 times
7 Same as Session 4 with op en and close rep eated 5 times
8 Op en-faced hand mo v es in one direction
9 Op en-faced hand mo v es in forw ard and bac kw ard motion rep eated 3 times
10 Op en-faced hand mo v es in left to righ t motion rep eated 3 times with elb o w in a xed lo cation
11 Hand squeezes a ball and then relaxes
12 Same as Session 11 with squeeze and release motion rep eated 3 times
13 Hand holds a ball and rotates ball coun terclo c kwise and then clo c kwise
14 P oin ter nger pushes ball in an up and do wn motion rep eated 4 times with hand in a closed-st
15 Hand squeezes a ball with th um b, p oin ter and middle ngers and then releases
16 All ngers excluding th um b pushes ball in an up and do wn motion rep eated 4 times
17 With an op en-faced hand, eac h nger starting with the pinky nger mo v es in a down andupmotion
rep eated 3 times
18 Same as Session 11 except squeezing and releasing action done with greater sp eed and force
T able 2: Recorded windo w sessions
5 Exp erimen tal Results
W e conducted sev eral exp erimen ts to: 1) study the impact of the condence threshold on the accuracy and eciency
of the acquisition, and 2) compare the accuracy and eciency of the three prop osed sampling approac hes (MFS, GS,
and AS). As a result of our studies, weha vealsoiden tied the optimal sampling rates for MFS and GS for Cyb erGrasp
devices.
F or our exp erimen ts, w e connected one Cyb erGrasp glo v e to a PC equipp ed with one P en tium I I 350 MHz CPU
and 128 MB of RAM, running Windo ws NT 4.0. The o wc hart in Figure 3 sho ws ho w the data from the Cyb erGrasp
device is pro cessed. A Butterw orth lo w-pass lter with degree 10 and adaptiv ely selected bandwidth (based on the
v alue of f
cut of f
) is used to the reduce the eect of the noise. The sinc function ( sin(x)=x ) is applied for up and
13
0
0.2
0.4
0.6
0.8
1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Window Session
Average Minimum Square Estimate (Error)
80%
90%
95%
97%
99%
Confidence
Threshold
0
20
40
60
80
100
1234 56789 101112131415161718
Window Session
Percentage of Reduction
90%
95%
97%
99%
Confidence
Threshold
a. Impact on error b. Impact on data and bandwidth reduction
Figure 4: The impact of condence threshold (CT) on data acquisition (p er session across all sensors)
Time
Value
Original Data Reconstructed Data, CT=90% Reconstructed Data, CT=95%
Time
Value
Original Data Reconstructed Data, CT=97% Reconstructed Data, CT=99%
a. Original and reconstructed data (CT=90%, 95% ) b. Original and reconstructed data (CT=97 %, 99%)
17 23
32
79
663
Data
Size (bytes)
Original Data
Reduced Data, CT=90%
Reduced Data, CT=95%
Reduced Data, CT=97%
Reduced Data, CT=99%
c. Size of the original and reduced size data
Figure 5: The impact of condence threshold on data acquisition (for sensor 5 within windo w session 7)
14
0
0.05
0.1
0.15
0.2
0.25
123456 789 10 11 12 13 14 15 16 17 18
Window Session
Data Size (MB)
Original Data Original Data-Compressed
Reduced Size Data Reduced Size Data-Compressed
Figure 6: Comparison b et w een data size for con v en tional compression tec hnique and adaptiv e sampling with CT =
90%
do wn sampling.
W e recorded 18 simple windo w sessions as describ ed in T able 2, eac h with 10 to 20 seconds duration. The memory
requiremen t for these windo w sessions is appro ximately 200 to 400 KB for the input data and 1 to 2 MB o v erall, and
the bandwidth requiremen t is 165 Kb/s. Our measuremen ts sho w ed that the complete pro cess of pro ducing reduced
size data from original data tak es b et w een 1.2 to 1.9 seconds, appro ximately 10% of the duration of the windo ws.
F or eachwindo w i and sensor j w e applied the basic sampling metho dology describ ed in Section 3.1 to compute the
sampling rate r
ij
. Figure 4 depicts the impact of CT on data acquisition. Wev aried CT from 80 to 99%, and for eac h
v alue of CT, w e computed eac h r
ij
. Subsequen tly , for eac h windo w session i,wea v eraged o v er the minim um square
estimate for eac h sensor to compute the accuracy of the windo w session. The reduced storage of windo w session i is
computed assuming that eac h sensor j, is sampled at the rate of r
ij
. As illustrated in Figure 4, the higher the CT the
more accurately the sampled data estimates the original data, and the higher its storage and bandwidth requiremen ts.
An in teresting observ ation is that CT=90% represen ts an acceptable accuracy while its storage/bandwidth sa ving is
signican tly b etter than higher CT s. Note that this observ ation is only true for Cyb erGrasp and the recorded windo w
sessions. More thorough exp erimen ts are required to study whether this observ ation is generalizable to haptic data
or immersidata. T o better visualize the impact of sampling the original data with dieren t CT s, w e compared the
reconstructed to the original data (in shap e and size). Figure 5 sho ws an example for sensor 5 within the windo w
session 7.
In Figure 6, w e xed CT at 90% and compared the storage reduction of adaptiv e sampling to the case where the
15
0
40
80
120
160
200
12 34567 89 10 11 12 13 14 15 16 17 18
Window Session
Required Bandwidth (kb/sec)
Fixed Sampling Modified Fixed Sampling Group Sampling Adaptive Sampling
Figure 7: Bandwidth requiremen ts for original data vs. MFS, GS and AS sampling tec hniques
No. Group Name Sensors Max. Sampling Rate (Hz) for the Group
1 Inner Join ts 2,5,8,12,16 35
2 Outer Join ts 3,7,10,14,18 18
3 Middle Join ts 6,9,13,17 29
4 Ab ductions 4,11,15,19 62
5 Lo cation 23,24,25 18
6 Rotation 26,27,28 62
7 F orces 29,30,31,32,33 67
Sensors in no group 1,20,21,22 8,35,13,8
T able 3: 7 dened groups and their maxim um rate across the 18 windo w sessions
windo w session is compressed b y standard Lemp el-Ziv approac h (e.g., Unix zip utilit y). The observ ation here is that
adaptiv e sampling can reduce the storage requiremen t signican tly b etter than a blac k-b o x compression tec hnique.
The zip utilit y reduces the size of the input data by4%to53% whilethe adaptiv e sampling pro duces 76% to 97%
reduction in storage and bandwidth. As sho wn in the gure, further compression on the reduced size data b y zip
utilit y resulted in to almost no storage sa vings (0.5 to 4%).
In Figure 7, w e compared the bandwidth requiremen t of our four prop osed sampling tec hniques. F or FS and MFS,
w e xed the sampling rate at 80 and 67 Hz, resp ectiv ely . F or GS, w e group ed the sensors in to 7 groups, as sho wn in
T able 3, and then used the maxim um rate for eac h group as demonstrated in Figure 8. Finally , for AS, w e used r
ij
’s
computed for CT=90%. As depicted in Figure 7, adaptiv e sampling requires far less bandwidth (and storage) compared
with other tec hniques, whic h is signican t enough to justify our future studies on designing a general framew ork for
adaptiv e sampling of dieren t sensors in an immersiveen vironmen t.
16
0
10
20
30
40
50
60
70
80
12 34 56 78 9 10 11 12 13 14 15 16 17 18
Window Session
Maximum Rate (Sample/Sec.)
Inner Joints Outer Joints Middle Joints Abductions
Locations Rotations Forces
Sensor
groups
Figure 8: Maxim um rate for the 7 groups dened in T able 3 across the 18 windo w sessions
6 Discussion
T o b etter understand our results, it is useful to rst classify sessions in immersiv een vironmen ts in to t w o groups: those
that require precise, real-time transmission of the data to m ultiple session participan ts and those that do not require
real-time transmission. The former case refers to a strict class of distributed whiteb oard st yle applications, where
not only are there m ultiple participan ts that aect eac h other in a virtual session, but where comm unication latency
bet w een the participan ts m ust be k ept to a minim um or be non-existen t. F or brevit y , w e shall refer to this sp ecial
case of immersiveen vironmentas r e al-time c o op er ative immersipr esenc e. Still, it is imp ortan t to note that there are
man y instances where m ultiple participan ts in a giv en virtual en vironmen t can tolerate latency (e.g., virtual lectures,
virtual ro oms where participan ts do not in teract with eac h other, etc.).
In reviewing our results on xed sampling, w e conclude that it is unquestionably the easiest approac h requiring no
additional implemen tation. Ho w ev er, it is also the most w asteful metho d, often sampling at rates that are unnecessarily
high. Despite its costs, xed sampling’s merit is that it can b e used for real-time co op erativ e immersipresence.
Mo died xed sampling is an optimization of xed sampling, with a (
r max r 0
r max
)% reduction in required bandwidth
and data storage (16% for our setup), but lik e xed sampling, its computational needs are minimal. One dra wbackto
this approac h is that it do es require initial training sessions in order to iden tify an optimal sampling rate for the device.
Another dra wbackis what the training session assumes: that the so-called optimal rate will suce in all scenarios.
This ma y not b e alw a ys true, since there are times when a high rate of sampling is necessary and other times when an
absolute minim um of sampling is required. Th us, while training sessions can iden tify a sampling rate that help reduce
17
the o v erall data (th us easing transmission and storage requiremen ts), it can cause the system to o v er- or under-sample
for b eha viors that are not captured in the training sessions. Lik e xed sampling, mo died xed sampling can also b e
used for real-time co op erativ e immersipresence.
Group ed sampling is a more in telligen t form of mo died xed sampling, since it targets groups of related sensors.
Our results sho w that group ed sampling led to further data size reductions of (
P
m
i=1
(r max r g
i
)
m:r max
)% (50% o v erall in our
setup), where m is n um b er of groups. As with xed sampling and mo died xed sampling, its computational needs
are minimal.While this is an impro v emen t o v er mo died xed sampling, it suers from the same p oten tial lac k of
sampling robustness and th us ma y lead to o v er- or under-sampling. In addition, group ed sampling requires a metho d
for lo cating the actual groups. In some cases, the groups ma y b e kno wn apriori (e.g., the Cyb erGrasp has nger force
sensors and nger angle sensors, whic h can b e seen as t w o distinct groups). Ho w ev er, in man y cases, it is necessary
to disco v er these groups and th us some sort of cluster disco v ery algorithm is required. Likethe other t w o metho ds,
group ed sampling requires training sessions, p erhaps more extensiv e ones in order to iden tify the groups, but can also
b e used for real-time co op erativ e immersipresence.
Finally,w e found that adaptiv e sampling, whic h con tin uously iden ties optimal sampling rates for eac h sensor on
a haptic device, w as the most ecien t metho d for acquiring haptic data. It substan tially reduced the data size b y
(
P
n
i=1
(r max r ij )
n:r max
)% for session j (76% to 97% in our scenarios). Ho w ev er, it is also the most complex to implemen t
b ecause it requires concurren t, p erio dic analysis of eac h sensor output for a giv en windo w of time b efore the optimal rate
for that sensor can b e iden tied. Th us, it do es ha v e some limitations in terms of real-time co op erativ e immersipresence.
Ho w ev er, w e again note that this is one mo de of immersidata acquisition - as stated earlier, other mo des (laten t
transmission, pure storage, etc) do not ha v e this demand. One w a y to reduce the latency in real-time co op erativ e
immersipresence is to adjust the windo w size (in eect, decrease the rate of adaptation). Finally , in addition to its
eciency ,av ery attractiv e feature of adaptiv e sampling is its robustness: o v er-sampling and under-sampling are not
problems with this approac h.
Our results can b e summarized in Figure 9. In particular, this graph illustrates the trade-os b et w een computational
needs and storage/net w ork eciency for eac hof the tec hniques in v estigated. F or xed, mo died xed and group ed
sampling, there is little computation required and they deliv er v ariable lev els of space eciency . While adaptiv e
sampling pro vides the most ecien t solution, it has more signican t computational requiremen ts.
18
Fixed Sampling
Modified Fixed
Sampling
Grouped Sampling
Adaptive Sampling
Computational Complexity
Bandwidth & Sotrage Requirements
Figure 9: T rade-os b et w een bandwidth/storage requiremen ts and computational complexit y .
7 Conclusion and F uture W ork
Weha v e iden tied sev eral metho ds for impro ving real-time sampling of haptic data. One of these approac hes, adap-
tiv e sampling, is particularly attractiv e b ecause of its eciency and robustness. In particular, the reduction in data
stored/transmitted p er second is substan tial, reducing bandwidth demands and th us enabling b etter scalabilit y for im-
mersiv e sessions or deplo ymento v er lo w er-bandwidth infrastructures. In addition, to handle the real-time co op erativ e
immersipresence case, w e can adjust adaptiv e sampling via t w o metho ds: (a) reduce the size of the windo w, so that
latency is decreased and/or (b) use mo died xed sampling for transmitting b et w een session participan ts but retain
adaptiv e sampling for database storage needs. One note with regards to the data reduction ac hiev ed b y adaptiv e
sampling is that additional compression (i.e,. Lemp el-Ziv metho ds) do es not further reduce the size of the data.
W e plan to con tin ue our w ork in v estigating the optimization of haptic data acquisition. One near-term goal w e
ha v e is to conduct more exp erimen ts so that w e can b etter understand the training sessions. In particular, wew ould
lik e to fo cus on the training windo w. Wew ould liketokno w what the optimal duration is and ho w comprehensivethe
actions p erformed need to b e in order to ac hiev e an acceptable lev el of accuracy . In eect, w ew ould lik e to b e able to
quan tify the p oten tial lac k of robustness when acquiring haptic data based on a training-generated sampling rate. In
the future, weare in terested in b eing able to extract seman tic information from the sampled haptic data. F or example,
if a user mak es a st for 20 seconds in a session, w ew ould lik e to b e able to extract that from the large amoun ts of data
normally sampled from that session. This has ob vious implications: if w e can extract seman tic information, wecan
drastically reduce the amoun t of data required to describ e a session, th us enabling us to more ecien tly transmit and
19
store haptic sessions. Ov erall, our con tin ual goal is to ac hiev e a b etter lev el of manageabilit y for immersidata suchas
haptic data. In the future, these en vironmen ts will lik ely consist of man y sensors and man y participan ts, placing ev en
higher demands on data acquisition. F or b oth real-time co op erativ e immersipresence and database storage purp oses,
it is imp ortan t to con tin ue researc h on ecien t metho ds that will address these an ticipated needs.
References
[AR99] W. A viles and J. Ran ta. A Brief Presen tation on the VRDTS - Virtual Realit y Den tal T raining System, In
Pr o c. F ourth PHANT oM Users Gr oup W orkshop, MIT, 1999.
[Mor98] Andrew B. Mor. 5 DOF F orce F eedbac k Using the 3 DOF PHANT oM and a 2 DOF Device. In Pr o c. PHANT oM
Users Gr oup, PUG98, 1998.
[KK97] Sang Hyun Kim and Nam Ch ul Kim, Lo w Bit Rate Video Co ding Using W a v elet-Based F ractal Appro xima-
tion, International Confer enc e on Image Pr o c essing, W ashington, DC, 1997.
[LHHC96] Hongc he Liu, T sai-Hong Hong, Martin Herman and Rama Chellappa, A ccuracy vs. Eciency T rade-os in
Optical Flo w Algorithms, Eur op e an Confer enc e on Computer Vision, Cambridge, UK, 1996.
[MSH+00] M. L. McLaughlin, G. Sukhatme, J. Hespanha, C. Shahabi, A. Ortega and G. Medioni. The Haptic Museum,
In Pr o c e e dings of EV A 2000 Confer enceonEle ctr onic Imaging and the Visual A rts, 2000.
[Nyq24] H. Nyquist. Certain F actors Aecting T elegraph Sp eed, Bel l System T e chnic al Journal , April 1924.
[PP99] W endy Plesniak and Ra vik an th P appu. Spatial In teraction With Haptic Holograms, In Pr o c e e dings of the IEEE
International Confer enc e on Multime dia Computing and Systems, June 1999.
[RDS94] F.L. Rob y , K. Dandek ar and M.A. Sriniv asan. Study of Fingertip Deformation Under Inden tations b y Cir-
cular and Rectangular Inden tors, R ep ort to the MIT Summer R ese ar ch Pr o gr am, Massachusetts Institute of T e chnol-
o gy,1994.
[RS95] F.L. Rob y and M.A. Sriniv asan. Study of Fingertip Deformations Under Inden tations b y Circular Inden tors,
R ep ort to the MIT Summer R ese ar ch Pr o gr am, Massachusetts Institute of T e chnolo gy,1995.
[SW48] C.E. Shannon and W. W ea v er. The Mathematical Theory of Comm unication, The University of Il linois Pr ess,
Urb ana, Il linois, 1948.
[SA V98] S.S. Snibb e, Anderson and B. V erplank. Springs and Constrain ts for 3D Dra wing, In Pr o c. PHANT oM Users
Gr oup, PUG98 , 1998
[SBE+99] Reference remo v ed for double blind reviewing.
[SGD92] M.A. Sriniv asan, R.J. Gulati and K. Dandek ar. In Viv o Compressibilit y of the Human Fingertip, In A dvanc es
in Bio engine ering, ASME Winter A nnual Me eting , No v ewm b er 1992.
20
[Sjo97] C. Sjostrom. The Phan tasticon: The PHANT oM for Disabled Children, Center of R ehabilitation Engine ering
R ese ar ch, Lund University. (1997) URL: h ttp:// www.certec.lth.se
[T an00] Hong Z. T an. Haptic In terfaces, Communic ations of the A CM, Marc h 2000.
21

Abstract (if available)

Linked assets

Computer Science Technical Report Archive

Conceptually similar

PDF

USC Computer Science Technical Reports, no. 742 (2001)

PDF

USC Computer Science Technical Reports, no. 799 (2003)

PDF

USC Computer Science Technical Reports, no. 785 (2003)

PDF

USC Computer Science Technical Reports, no. 733 (2000)

PDF

USC Computer Science Technical Reports, no. 840 (2005)

PDF

USC Computer Science Technical Reports, no. 748 (2001)

PDF

USC Computer Science Technical Reports, no. 694 (1999)

PDF

USC Computer Science Technical Reports, no. 736 (2000)

PDF

USC Computer Science Technical Reports, no. 855 (2005)

PDF

USC Computer Science Technical Reports, no. 870 (2005)

PDF

USC Computer Science Technical Reports, no. 766 (2002)

PDF

USC Computer Science Technical Reports, no. 740 (2001)

PDF

USC Computer Science Technical Reports, no. 839 (2004)

PDF

USC Computer Science Technical Reports, no. 713 (1999)

PDF

USC Computer Science Technical Reports, no. 964 (2016)

PDF

USC Computer Science Technical Reports, no. 968 (2016)

PDF

USC Computer Science Technical Reports, no. 659 (1997)

PDF

USC Computer Science Technical Reports, no. 826 (2004)

PDF

USC Computer Science Technical Reports, no. 966 (2016)

PDF

USC Computer Science Technical Reports, no. 650 (1997)

Description

Cyrus Shahabi, Greg Barish, Mohammad R. Kolahdouzan, Didi Yao, Roger Zimmermann, Kun Fu, Lingling Zhang. "Alternative techniques for efficient acquisition of haptic data." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 739 (2001).

Asset Metadata

Creator Barish, Greg (author), Fu, Kun (author), Kolahdouzan, Mohammad R. (author), Shahabi, Cyrus (author), Yao, Didi (author), Zhang, Lingling (author), Zimmermann, Roger (author)

Core Title USC Computer Science Technical Reports, no. 739 (2001)

Alternative Title Alternative techniques for efficient acquisition of haptic data (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 21 pages (extent), technical reports (aat)

Language English

Unique identifier UC16269114

Identifier 01-739 Alternative Techniques for Efficient acquisition of Haptic data (filename)

Legacy Identifier usc-cstr-01-739

Format 21 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

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)