Li S.Z., Jain A.K. (eds.) Encyclopedia of Biometrics

Подождите немного. Документ загружается.

enrollment, and matching failures. However, further

understanding in the area of biometric system ergo-

nomic design and its impact on biometrics is needed

to meet this goal.

The authors are not alone in their thoughts

and opinions that continued research is needed in

the area of biometric system ergonomic design. As

Smith [11] stated that some members of the HCI

community believe that interfaces of security systems

do not reﬂect good thinking in terms of creating a

system that is easy to use, while maintaining an accept-

able level of security. Moreover Adams and Sasse dis-

cussed the fact that security systems are one of the last

areas to embrace user-centered design and training

as essential [12]. Lastly, Maple and Norrington [20],

noted three observations that align with the objective

for continued investigation in biometric system ergo-

nomic design:



People have different cognitive abilities,



People have different physical characteristics and

interact differently with equipment, and



People have different sensory abilities and will per-

ceive biometric sensors and systems differently.

As the biometrics community continues to develop

biometric systems and deployments become more per-

vasive, the evaluation of the biometric system and the

respective human-biometric sensor interaction w ill

continue to gain traction.

Related Entries

▶ Accessibility

▶ Attempt

▶ Failure to Acquire (FTA)

▶ Failure to Enroll (FTE)

▶ Usuability

References

1. International standards organization, information technology –

Biometric performance testing and reporting – Part 1: Principles

and framework. ISO/IEC: Geneva. p. 56. (2006)

2. International standards organization, information technology –

Biometric performance testing and reporting - Part 2: Testing

methodologies for technology and scenario evaluation. ISO/IEC:

Geneva. p. 48. (2007)

3. International standards organization, Text of DTR 19795-3,

Biometric performance testing and reporting – Part 3:

Modality-speciﬁc testing. ISO/IEC: Geneva. p. 28. (2007)

4. Clarke, R.: Human identiﬁcation in information systems: Man-

agement challenges and public policy issues. Inf. Technol. People

7(4), 6–37 (1994)

5. Young, M., Elliott, S.: Image Quality and Performance Based on

Henry Classiﬁcation and Finger Location. In IEEE Workshop on

Automatic Identiﬁcation Advanced Technologies. Alghero, Italy

(2007)

6. Wayman, J.: Multi-ﬁnger penetration rate and ROC variability

for automatic ﬁngerprint identiﬁcation systems. In: Wayman, J.

(ed.) National Biometric Test Center Collected Works

1997–2000, pp. 179–190. San Jose, CA (2000)

7. Kukula, E.: Design and Evaluation of the Human-Bio-

metric Sensor Interaction Method. In: Industrial Tech-

nology, Vol. 1, Ph.D, pp. 510. Purdue University: West

Lafayette (2008)

8. Kukula, E., Elliott, S., Duffy, V.: The effects of human interaction

on biometric system performance in 12th International Confer-

ence on Human-Computer Interaction and 1st International

Conference on Digital-Human Modeling. pp. 903–913, Springer,

Beijing, China (2007)

9. Yao, M., Pankanti, S., Haas, N.: Fingerprint quality assessment.

In: Ratha, N., Bolle, R. (eds.) Automatic Fingerprint Recognition

Systems, pp. 55–66. Springer, New York (2004)

10. Rubin, R.: Handbook of usability testing: How to plan, design,

and conduct effective tests. Wiley, New York (1994)

11. Smith, S.: Humans in the loop: Human-computer interaction

and security. IEEE Secur. Priv. 1(3), 75–79 (2003)

12. Adams, A., Sasse, M.: Users are not the enemy : Why users

compromise security mechanisms and how to take remedial

measures. Commun. ACM 42(12), 41–46 (1999)

13. Coventry, L., De Angeli, A., Johnson, G.: Usability and biometric

veriﬁcation at the ATM interface. In: Conference on Human

Factors in Computing Systems. Ft. Lauderdale, ACM Press,

Florida (2003)

14. Tayyari, F., Smith, J.: Occupational ergonomics: Principles and

applications. In: Parsaei, H. (ed.) Manufacturing Systems Engi-

neering Series, p. 452. Kluwer Academic Publishers, Norwell

(2003)

15. International organization for standardization, ISO 9241: Ergo-

nomic requirements for ofﬁce work with visual display terminals

(VDTs) - Part 11: Guidance on usability. p. 28. (1998)

16. Coventry, L.: Usable biometrics. In: Cranor, L.F., Garﬁnkel, S.

(eds.) Security and usability: Designing secure systems that

people can use, pp. 175–198. O’Reilly Media, Inc: Sebastopol,

CA (2005)

17. NIST. Biometrics and usability group. Available from: http://

zing.ncsl.nist.gov/biousa/index.html (2007). Accessed 30

November 2007

18. Purdue University Biometric Standards Performance & Assur-

ance Laboratory. Human Biometric Sensor Interaction. 2007

[cited 2007 November 30]; Available from: http://www.bspa

labs.org/archives/category/research/hbsi

19. Home ofﬁce identity & passport service. Publications. Available

from: http://www.ips.gov.uk/passport/publications-general.asp

(2007). Accessed 30 November 2007

20. Maple, C., Norrington, P.: The usability and practicality of

biometric authentication in the workplace. In: First Interna-

tional Conference on Availability, Reliability and Security

(ARES’06). IEEE, Vienna, Austria (2006)

280

Ergonomic Design for Biometric Systems

Ergonomics

Ergonomics is a derivative of the Greek words ‘‘ergon,’’

or work, and ‘‘nomos,’’ meaning laws. While the term

work has been traditionally associated with occupation,

a broader sense of the term can be applied to any

unplanned activity requiring skill or effort. In 2000,

the International Ergonomics Association (IEA) de-

ﬁned ergonomics or human factors as: ‘‘The scientiﬁc

discipline concerned with the understanding of inter-

actions among humans and other elements of a system,

and the profession that applies theory, principles, data

and methods to design in order to optimize human

well-being and overall system performance.’’ In design,

ergonomics attempts to achieve an optimal relation-

ship between humans and machines in a particular

environment. The goal of ergonomics, according to

Tayyari and Smith, is to ‘‘ﬁt (adapt) work to indivi-

duals, as opposed to ﬁtting individuals to the work.’’

▶ Ergonomic Design for Biometric System

Error Probability Non-Accumulation

▶ Score Normalization Rules in Iris Recognition

Evaluation of Gait Recognition

SUDEEP SARKAR

,ZONGYI LIU

Computer Science and Engineering, University of

South Florida, Tampa, FL, USA

Amazon.com, Seattle, WA, USA

Synonyms

Gait recognition

Definition

Gait recognition refers to automated vision methods

that use video of human gait to recognize or to identify

a person. Evaluation of gait recognition refers to the

benchmarking of progress in the desig n of gait recog-

nition algorithms on standard, common, datasets.

Introduction

Design of biometric algorithms and evaluation of

performance goes hand in hand. It is important to

constantly evaluate and analyze progress being at vari-

ous levels of biometrics design. This evaluation can

be of three types: at algorithm-level, at scenario-level,

and at operational-level, roughly corresponding to the

maturity of the biometric. Given the young nature

of gait as a biometric source, relative to the mature

biometrics such as ﬁngerprints, current evaluations are

necessarily at algorithm-level. The motivation behind

algorithm-level evaluations is to explore possibilities,

to understand limitations, and to push algorithmic re-

search towards hard problems. Some of the relevant

questions are

1. Is progress being made in gait recognition of

humans?

2. To what extent does gait offer potential as an iden-

tifying biometric?

3. What factors affect gait recognition and to what

extent?

4. What are the critical vision components affecting

gait recognition from video?

5. What are the strengths and weaknesses of different

gait recognition algorithms?

An overview of the current evaluation of gait as a

potential biometric is discussed here, with particular

emphasis on the progress with respect to the HumanID

gait challenge problem that has become the de-facto

benchmark. A synthesis of gait recognition perfor-

mances reported on this dataset and other major

ones is provided here, along with some suggestions

for future evaluations.

A Panoramic View of Performance

To take stock of the progress made in gait recognition,

consider a summary of the identiﬁcation rates reported

in the recent literature on different kinds of publicly

available experimental protocols and datasets (>25

persons) such as the CMU-Mobo dataset [1] (indoor,

Evaluation of Gait Recognition

281

25 subjects), the UMD dataset [2] (outdoor, 55 sub-

jects), the Southampton Large dataset [3] (indoor and

outdoor, 115 subjects), the CASIA Gait Dataset [4]

(indoor, 124 subjects), and the HumanID Gait Chal-

lenge dataset [5] (outdoor, 122 subjects). Figure 1 lists

the average identiﬁcation rates for matching across

different conditions, i.e., the

▶ probe and the gallery

differed with respect to the indicated

▶ covariate.Of

course, the caveat is that the conclusions are conditioned

on the kinds of variations of each covariate observed in

the respective datasets. Hence, a deﬁnitive conclusion

is hard to make. However, this kind of summary has

some conclusive weight since, it encompasses the ﬁnd-

ings of multiple research groups. It should provide

some directions for focusing future research. The

data shows that outdoor gait recognition, recognition

across walking surface-type change, and recognition

across months are all hard problems. Clothing, foot-

wear, carrying condition, and walking speed does not

seem to be hard covariates to overcome. As expected,

performance also drops with dataset size, which sug-

gests that it is imperative to demonstrate the efﬁcacy of

an idea on as large a dataset as possible.

A deeper look at the performances reported

on commonly available datasets, in particular the

HumanID gait challenge problem, will form the basis

for more deﬁnitive conclusions about the progress that

is being made.

The HumanID Gait Challenge Problem

The development of gait biometrics is following a path

that is somewhat different from other biometrics, for

which serious evaluation benchmarks appeared after

years of algorithmic development. It was more than

20 years for face recognition, whereas evaluation

framework for gait recognition appeared in less than

10 years after the ﬁrst publication of vision algorithms

for gait recognition. Bulk of the research in gait recog-

nition was spurred by the DARPA HumanID at a

distance program. The HumanID gait challenge prob-

lem was formulated in this program to facilitate objec-

tive, quantitative measurement of gait research

progress on a large dataset [5]. As of end of 2007, this

dataset has been distributed to more than 40 research

groups. Many gait recognition research papers report

performance on this dataset.

This challenge problem does not just consist of a

dataset, but also provides a well-deﬁned experimental

Evaluation of Gait Recognition. Figure 1 Summary (average) of gait identification rates as reported in the literature

for different conditions. The average of the reported rates are listed for different conditions. The first two performance

points are average of reported rates on datasets collected indoors and outdoors, respectively. The next six performance

points are for matching gait templates across different conditions. For example, ‘‘carrying condition’’ refers to matching

gait sequences where the hands of the subjects were free to sequences where the subjects were carrying a briefcase. The

‘‘time’’ condition refers to matching gait templates collected at different times with the time-gap as noted. The size

condition refers to number of subjects used in the experiments.

282

Evaluation of Gait Recognition

framework for others to follow, along with an estab-

lished benchmark.

The Dataset

The data was collected outdoors. For each person in

the data set, there are combination of as many as ﬁve

conditions or covariates. The conditions are: (1) two

camera angles (L and R), (2) two shoe types (A and B),

(3) two surfaces (grass and concrete), (4) with and

without carrying a briefcase (B or NB), and (5) two

different dates 6 months apar t, May and November.

The covariates were chosen based on consultation with

gait recognition researchers in the Human ID program.

These are, of course, not the only variables that

can impact gait , but were logistically feasible and likely

to impact gait the most. Attempt was made to acquire a

person’s gait in all possible combinations, and there are

up to 32 sequences for some persons. Hence, the full

dataset can be partitioned into 32 subsets, one for each

combination of the ﬁve covariates. The partitioning of

the data is visualized in Fig. 2. Each cell refers to an

unique combination of view, shoe type, and surface

covariates. The smaller arrangement of cells represent

the data from repeat subjects. Comparisons between

these subsets are used to set up challenge experiments;

more on this later. The full data set consists of 1,870

sequences from 122 individuals. This dataset is unique

in the number of covariates exercised. It is the only

data set to include walking on a grass surface. Figure 3

shows some sample frames from this dataset.

In addition to the raw data sequence, there is an

ancillary information associated with the data. First, for

each sequence, there is meta-data information about the

subject’s age, sex, reported height, self reported weight,

foot dominance, and shoe information.

Second, for a subset of this dataset, manually created

▶ silhouettes (see Fig . 4) are available. These manual

silhouettes should not be used to test any recogni-

tion algorithm, but they could be used to build models

or to study segmentation errors. More details about

the process of creating these manual silhouettes

and the quality checks performed can be found in

[6]; here are some salient aspects. Seventy one subjects

from one of the two collection periods (May collection)

were chosen for manual silhouette speciﬁcation. The

sequences corr esponding to these subjects were chosen

from the (1) gallery set (sequences tak en on grass, with

shoe type A, right camera view), (2) probe B (on grass,

with shoe type B, right camera view), (3) probe D (on

concrete, with shoe type A, right camera view), (4) probe

H (on grass, with shoe A, right camera view, carry-

ing briefcase), and probe K (on grass, elapsed time).

The silhouette in each frame over one walking cycle,

of approximately 30–40 image frames was manually

Evaluation of Gait Recognition. Figure 2 Partitioning of the HumanID gait challenge dataset in terms of its

covariates, which are coded as follows: C – concrete surface, G – grass surface, A – first shoe type, B – second shoe type,

BF–carryingabriefcase, NB–nobriefcase, M–datacollectedinMay,N

– new subjects in November data, and N

– repeat

subjects in November. The shaded cells are used to design the challenge experiments.

Evaluation of Gait Recognition

283

speciﬁed. This cycle was chosen to begin at the right heel

strike phase of the walking cycle through to the next

right heel strike. Whenever possible, this gait cycle was

selected from the same 3D location in each sequence. In

addition to marking a pixel as being from the back-

ground or subject, more detailed speciﬁcations in

terms of body parts were mark ed. The head, torso, left

arm, right arm, left upper leg, left lower leg, right upper

leg, and right lower leg were explicitly labeled using

different colors.

The Challenge Experiments

Along with the dataset, the gait challenge problem

includes a deﬁnition of 12 challenge experiments (A–L),

spanning different levels of difﬁculty. This provides a

common benchmark to compare performance with

other algorithms. The experiments are designed to

investigate the effect on performance of ﬁve factors,

i.e., change in viewing angle, change in shoe type,

change in walking surfaces (concrete and grass),

Evaluation of Gait Recognition. Figure 3 Frames from (a) the left camera for concrete surface, (b) the right camera for

concrete surface, (c) the left camera for grass surface, (d) the right camera for grass surface.

Evaluation of Gait Recognition. Figure 4 Top row shows the color images, cropped around the person, for one

sequence. The bottom row shows the corresponding part-level, manually speciﬁed silhouettes.

284

Evaluation of Gait Recognition

carrying or not carrying a briefcase, and temporal

differences. The gallery set is common for all the

experiments and corresponds to the dark colored cell

in Fig. 2. The gallery consists of sequences with the

following covariates: Grass, Shoe Type A, Right Cam -

era, No Briefcase, and collected in May along with

those from the new subjects from November. This set

was selected as the gallery because it was one of the

largest for a given set of covariates. The experiments

differ in terms of the probe sets, which are denoted by

the lightly shaded cells. The structure of the 12 probe

sets is listed in Table 1. The signatures are the video

sequences of gait. The last two experiments study the

impact of elapsed time. The elapsed time covariate

implicitly includes a change of shoe and clothing

because the subjects were not required to wear the

same clothes or shoes in both data collections. Because

of the implicit change of shoe, it can be safely assumed

that a different set of shoes were used in the May and

November data collections. This is noted in Table 1 by

A/B for shoe type in experi ments K and L. The key

experiments are those that involve controlled change in

just one covariate and are marked with an asterisk in

the table. The results from the 12 experiments provide

an ordering of difﬁculty of the experiments.

Baseline Gait Algorithm

The third aspect of the gait challenge problem is a

simple but effective

▶ baseline algorithm to provide

performance benchmarks for the experiments. Ideally,

this should be a combination of ‘‘standard’’ vision mod-

ules that accomplishes the task. Drawing from the suc-

cess of template based recognition strategies in

computer vision, a four-part algorithm that relies on

silhouette template matching was designed. The ﬁrst

part semi-automatically deﬁnes bounding boxes around

the moving person in each frame of a sequence. The

second part extracts silhouettes from the bounding

boxes using expectation maximization based on Maha-

lanobis distance between foreground and background

color model at each pixel. Each silhouette is scaled to a

height of 128 pixels and centered (automatically) in

each frame along the horizontal direction so that the

centerline of the torso is at the middle of the frame.

Evaluation of Gait Recognition. Table 1 The gallery and probe set specifications for each of gait challenge experiments

Probe

Exp.

(surface, shoe, view, carry, elapsed

time) (C/G, A/B, L/R, NB/BF, time) Number of subjects Difference

(G, A, L, NB, M + N

) 122 V

(G, B, R, NB, M + N

)54S

C (G, B, L, NB, M + N

) 54 S+V

(C, A, R, NB, M + N

) 121 F

E (C, B, R, NB, M + N

) 60 F+S

F (C, A, L, NB, M + N

) 121 F+V

G (C, B, L, NB, M + N

) 60 F+S+V

(G, A, R, BF, M + N

) 120 B

I (G, B, R, BF, M + N

) 60 S+B

J (G, A, L, BF, M + N

) 120 V+B

(G, A/B, R, NB, N

)33T

+S+C

L (C, A/B, R, NB, N

) 33 F+T+S+C

The gallery for all of the experiments is (G, A, R, NB, M +N

) and consists of 122 individuals

Keyexperiments

View

Shoe

Surface

Carry

Elapsed time

Clothing

Evaluation of Gait Recognition

285

The third part computes the gait period from the sil-

houettes. The gait period is used to partition the

sequences for spatial-temporal correlation. The fourth

part performs spatial-temporal correlation to compute

the similarity between two gait sequences.

Let S

¼fS

ð1Þ; ; S

ðMÞgand S

¼fS

ð1Þ;

; S

ðNÞg, be the probe and the gallery silhouette

sequences, respectively. First, the probe (input) se-

quence is partitioned into subsequences, each roughly

over one gait period, N

Gait

. Gait periodicity is esti-

mated based on periodic variation of the count the

number of foreground pixels in the lower part of the

silhouette in each frame over time. This number will

reach a max imum when the two legs are farthest apart

(full stride stance) and drop to a minimum when the

legs overlap (heels together stance).

Second, each of these probe subsequences, S

ðkÞ; ; S

ðk þ N

Gait

Þg, is cross correlated with the

given gallery sequence, S

CorrðS

; S

ÞðlÞ¼

Gait

j¼1

S S

ðk þ jÞ; S

ðl þ jÞðÞ; ð1Þ

where, the similarity between two image frames,

S(S

(i), S

(j)), is deﬁned to be the Tanimoto similarity

between the silhouettes, i.e., the ratio of the number of

common pixels to the number of pixels in their uni on.

The overall similarity measure is chosen to be the

median value of the maximum correlation of the gal-

lery sequence with each of these probe subsequences.

The strategy for breaking up the probe sequence into

subsequences allows the algorithm to overcome seg-

mentation errors in some contiguous sets of frames

due to some background subtraction artifact or due to

localized motion in the background.

SimðS

; S

Þ¼Median

max

CorrðS

; S

ÞðlÞ



: ð2Þ

The baseline algorithm is parameter free. The algo-

rithm, although straightforward, performs quite well

on some of the experiments and is quite competitive

with the ﬁrst generation of gait recognition algorithms.

Performance on the Gait Challenge

Problem

The results reported for the Gait Challenge problem

are of two types, ones that report results on the

ﬁrst version of the dataset that was released with 71

subjects and the second set of results are those reported

for the full dataset with 122 subjects. The smaller

dataset allows just the ﬁrst eight experiments listed in

Table 1, but with reduced gallery set sizes. Figure 5a

tracks the baseline performance and the best perfor-

mance reported in the literature. As of middle of 2007 ,

there were 18 papers that reported results on the smal-

ler version of the problem. In 2002, when the Gait

Challenge Problem was released, the performance of

the baseline algorithm was better than the best reported

performance. By 2004, whi le the baseline algorithm

performance improved as the algorithm was ﬁne-

tuned, the performance of the best performance im-

proved signiﬁcantly and continued to improve through

2006. This trend is also seen in the results reported in

six papers on the full dataset, summarized in Fig. 5b.

As is evident, the gait challenge problem has al-

ready spurred the development of gait recognition

algorithms with improving performanc e. What is partic-

ularly interesting to notice is that the performance

on har d experiments such those across surfac e (experi-

ment D) and elapsed time (experiment K) has improved.

Of course, there is still room for further impro vement.

Another interesting aspect is that the improvement of

performance from 2004 to 2006 was not due to

‘‘ c ontinue d engineering’’ of existing approaches, but

inv olv ed the redesign of the recognition appr oaches.

For instanc e, the greatest gains came from appr oaches

that analyzed the silhouette shapes rather than the dy-

namics [2, 7]. Dynamics is important, but by itself is

not sufﬁcient.

Performance of a large number of algorithmic

approaches have been explored. A review of the per-

formances reported in these works reveals* [8] that

1. All most all of these approaches are based on

silhouettes.

2. There is no one method that performs the best on

all experiments.

3. Performances that involve matching against view-

point and shoe variations, but on the same walking

surface, has plateaued out.

4. Matching against walking surface variation remains

a challenge.

5. Good performances (>80%) has been reported for

matching w ith and without carrying objects.

6. Matching across 6 months time difference has low

performance, but the number of subjects involved

in this experiment (33 subjects) is too low to derive

meaningful conclusions.

286

Evaluation of Gait Recognition

Other Large Datasets

There are currently two other datasets that are as large as

the H umanID gait challenge dataset in terms of number

of subjects. First is the CASIA Indoor Gait Database [4].

The gallery set includes 124 subjects with normal walk-

ing, no coat and no carr y bag. Different probes can be

deﬁned in terms of changes in (1) viewpoint, (2) cloth-

ing change (coat vs. no-coat), and (3) carrying a bag

and not carrying a bag. Not many algorithms have

Evaluation of Gait Recognition. Figure 5 Improvement in gait recognition algorithms over time with respect to the

baseline performance. (a) Results on the ﬁrst release of the gait dataset with 71 subjects in the gallery for the ﬁrst eight

experiments listed in Table 1 are tracked here. (b) Results on the full dataset with 122 subjects for the key experiments

listed in Table 1. From 2004 to 2006, the best reported performances are better on all the experiments.

Evaluation of Gait Recognition

287

reported performance on this dataset yet. But, the

performance reported for the gait energy image ap-

proach in [9], seems to corroborate the ﬁndings from

the HumanID Problem for matching across carrying

conditions: a performance of upto 80% is reported for

the CASIA dataset.

The other large dataset is the SOTON HID Gait

Database [10] with 115 subjects are collected mostly

indoor and some under outdoor conditions. The in-

door SOTON datase t was collected to examine the

premise that gait is unique so the background is

controlled so as to allow easy segmentation. The

same subjects were also ﬁlmed walking outdoors

to determine whether gait biometrics could be per-

ceived with complex backgrounds. Performances in

the range of 72–85% have been reported for matching

across sessions using a variety of approaches. This

dataset also affords the matching across time issue. It

has been shown that a time-dependent predictive

model [11 ] results in 92% recognition, but only for

ten subjects.

It is worth noting that face recognition on

these data sets would be poor, indeed given the low-

resolution and the uncontrolled lighting.

Future Evaluations

It is to be expected that each gait research group

would collect their own data set to develop ideas.

This is an important process. For instance, one new

dataset is the CASIA infrared night gait dataset [12].

It consists of gait data from 153 subjects are collected

outdoors, at night, with and without carrying condi-

tion, and at two different speeds. This dataset nicely

complements existing datasets that are collected dur-

ing the day. Given the data-driven nature of bio-

metrics research, the key to future progress are such

data sets, collected to explore issues not considered or

raised by existing ones. For instance, as of today there

is need for the better understanding of the variation

of gait due to surface conditions and across elapsed

time. Also, currently there is no dataset to explore the

matching across time issue for a large number of

subjects.

Ideally, the new datasets should consist of gait data

from around 1,000 subjects, an order of magnitude

larger than current large datasets. It is important to

increase the number of subjects so, that it is possible to

empirically study the scaling of perf ormance with

number of subjects. Some guidance about the requi-

red sizes can be found in [13, 14], where statistical

reasoning is employed to relate the number of subjects

with target error conﬁdences. The data collection

should include gait data repeated at regular time inter-

vals of weeks, spanning about a year. The datase t

should be collected in outdoor conditions, preferably

collected at a distance of 300 m to reﬂect real

world conditions. The dataset should come with a set

of well deﬁned experiments in terms of gallery

and probe sets. These experiments will inﬂuence the

types of algorithms. For the experiments to be effective

at inﬂuencing the direction of gait research the design

of the experiments needs to solve the three bears prob-

lem; the experiments must be neither too hard nor

too easy, but just right. If performance on the experi-

ments is easily saturated, then the gait recognition

community will not be challenged. If experiments are

too hard, then it will not be possible to make progress

on gait recognition. Ideally, the set of experiments

should vary in difﬁculty, characterize where the gait

recognition problem is solvable, and explore the fac-

tors that affect performance. A set of experiments

cannot meet this ideal unless the appropriate set

of data is collected. It is important to view biometrics

research as a data-driven algorithm development

process rather than algorithm-driven data collection

process.

Related Entries

▶ Hidden Markov Models

▶ Human Detection and Tracking

▶ Human Movement, Psychology

▶ Performance Evaluation, Overview

▶ Performance Testing Methodology Standardization

▶ Psychology of Gait Recognition

▶ Silhouette-B ased Gait and Action Recognition

▶ Veriﬁcation/Authentication/Identiﬁcation/

Recognition

References

1. Gross, R., Shi, J.: The CMU motion of body (MoBo) database.

Tech. report CMU-RI-TR-01-18, Robotics Institute, Carnegie

Mellon University (2001)

288

Evaluation of Gait Recognition

2. Kale, A., Sundaresan, A., Rajagopalan, A.N., Cuntoor, N.P.,

Roy-Chowdhury, A.K., Kru

ger, V., Chellappa, R.: Identiﬁ-

cation of humans using gait. IEEE Transactions on Image Pro-

cessing 13(9), 1163–1173 (2004)

3. Wagg, D.K., Nixon, M.S.: On automated model-based extraction

and analysis of gait. In: International Conference on Automatic

Face and Gesture Recognition, pp. 11–16 (2004)

4. Yu, S., Tan, D., Tan, T.: A framework for evaluating the effect of

view angle, clothing and carrying condition on gait recognition.

In: International Conference on Pattern Recognition, vol. 4,

pp. 441–444 (2006)

5. Sarkar, S., Jonathon Phillips, P., Liu, Z., Robledo-Vega, I.,

Grother, P., Bowyer, K.W.: The Human ID gait challenge prob-

lem: Data sets, performance, and analysis. IEEE Transactions on

Pattern Analysis and Machine Intelligence 27(2), 162–177

(2005)

6. Liu, Z., Malave, L., Osuntogun, A., Sudhakar, P., Sarkar, S.:

Toward understanding the limits of gait recognition. In: Proc.

of SPIE Defense and Security Symposium: Biometric Tech-

nology for Human Identiﬁcation, pp. 195–205 (2004)

7. Liu, Z., Sarkar, S.: Improved Gait Recognition by Gait Dynamics

Normalization. IEEE Transactions on Pattern Analysis and Ma-

chine Intelligence 28(6), 863–876 (2006)

8. Sarkar, S., Liu, Z.: Handbook of Biometrics, chap. Gait Recogni-

tion. Springer (2008)

9. Han, J., Bhanu, B.: Statistical feature fusion for gait-based

human recognition. In: IEEE Conference on Computer Vision

and Pattern Recognition, vol. II, pp. 842–847 (2004)

10. Shutler, J.D., Grant, M.G., Nixon, M.S., Carter, J.N.: On a

large sequence-based human gait database. In: International

Conference on Recent Advances in Soft Computing, pp. 66–71

(2002)

11. Veres, G., Nixon, M., Carter, J.: Model-based approaches for

predicting gait changes over time. In: International Conference

on Intelligent Sensors, Sensor Networks and Information,

pp. 325–330 (2005)

12. Tan, D., Huang, K., Yu, S., Tan, T.: Efﬁcient night gait recogni-

tion based on template matching. In: International Conference

on Pattern Recognition, vol. 3, pp. 1000–1003 (2006)

13. Doddington, G., Przybocki, M., Martin, A., Reynolds, D.: The

NIST speaker recognition evaluation–Overview, methodology,

systems, results, perspective. Speech Communication 31(2–3),

225–254 (2000)

14. Veres, G., Nixon, M., Carter, J.: Is Enough Enough?? What Is

Sufﬁciency in Biometric Data?. Lecture Notes in Computer

Science 4142, 262 (2006)

Expected Performance or Utility of

Fingerprint Image in an Automated

Comparison Environment

▶ Fingerprint Image Quality

Expression

Face recognition processing usually requires a neutral

facial expression (no smiling, mouth closed) and with

the subject looking straight at the camera.

▶ Photography for Face Image Data

Extended Enterprise

Beyond the physical boundaries or ‘‘four walls’’ of an

organization is extended enterprise.

▶ Remote Authentication

External Identification

Identiﬁcation of a victim based on external evidence,

such as the victim’s gender, height, build, face, and

ﬁngerprints.

▶ Dental Biometrics

External Operation Time

▶ Operational Times

Extra-Class

Extra-class refers to instances of different subjects.

Ideally, the extracted features should be very different

for instances of different subjects.

▶ Local Image Filters

Extra-Class

289