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

Подождите немного. Документ загружается.

Jing et al. [13] propose a sensor level fusion algo-

rithm that generates a composite image from face and

palmprint biometrics. Circular Gabor ﬁlters are ﬁrst

applied on face and palm print images to generate 32

ﬁltered responses of each biometric data. These ﬁltered

responses are concatenated to generate a fused image.

A pixel normalization scheme is then used to minimize

variations due to imaging conditions. Finally, kernel

discriminative common vectors are extracted from the

fused image and radial basis function based neural net-

work is used for classiﬁcation. The fusion algorithm

improves the recognition performance and is an effective

solution for the small sample size recognition problem.

Noore et al. [14] proposed discrete wavelet transforma-

tion based image fusion algorithm that generates

a composite image by combining multim odal bio-

metric images. The algorithm starts with transforming

biometric images into wavelet domain and genera-

ting composite image by amalgamating the wavelet

coefﬁcients. The composite image is then scrambled

using a secret encoding key generated with Fibonacci

transforms. The algorithm not only improves the

recognition performance but also reduces the memory

requirements and provides resilience to common

image processing attacks such as smoothing, cropping,

JPEG 2000 compression, and ﬁltering.

Future Research Directions

As discussed in previous sections, sensor level fusion

has several advantages. However, compared to other

levels of fusion, this level of fusion is less explored and

requires fur ther research to address the limitations of

current research. First and foremost is to further im-

prove the recognition accuracy. Researchers have

shown that for certain applications, sensor level fusion

algorithms do not provide better results compared to

match score level fusion algorithms [5, 11]. This is

mainly because existing algorithms do not effectively

reconcile the information that is useful for recognition.

We believe that existing sensor level fusion algorithms

fail in some cases because during information fusion

it is possible that redundant and less discriminatory

features become predominant. Furthermore, there is

a lack of generalized sensor level fusion algorithms

that can be used for different biometric scenarios or

applications. For instance, genetic algorithm based

multi-spectral image fusion algorithm can not be

directly used for multimodal image fusion. Additional

research is required to design an effective and generalized

sensor level fusion algorithm which can be applied to

different biometric modalities. Every sensor level fusion

algorithm requires speciﬁc feature extraction algorithm

that can effectively extract discriminatory biometric in-

formation from the composite image or data. This

requirement is not mandatory with a generalized sen-

sor level fusion algorithm. Therefore, a generalized al-

gorithm can be easily incorporated in commercial

systems and can conform to data fusion standards.

Another important research issue is to unify the

sensor level fusion in a

▶ uniﬁcation framework that

reconciles multiple fusion algorithms. Originally pro-

posed by Vatsa et al. [15], a biometric uniﬁcation

framework combines multiple fusion algorithms by

dynamically selecting the most appropriate fusion algo-

rithm depending on the input evidences such as quality

and other priors. Currently, the uniﬁcation framework

includes only the match score fusion algorithms. How-

ever, with proper modiﬁcations, the uniﬁcation frame-

work can be expanded to include multi-level fusion

algorithms that can address the operational needs

of biometric systems and provide better recognition

performance.

Related Entries

▶ Data Fusion

▶ Face Recognition

Fusion, Sensor-Level. Figure 4 Multimodal image fusion. (a) Image pertaining to different biometric modalities and

(b) Fused and scrambled image.

620

Fusion, Sensor-Level

▶ Identiﬁcation

▶ Veriﬁcation

References

1. Ross, A., Nandakumar, K., Jain, A.K.: Handbook of Multibio-

metrics. 1st edn. Springer, New York (2006)

2. Singh, R., Vatsa, M., Ross, A., Noore, A.: A mosaicing scheme for

pose-invariant face recognition. IEEE Trans. Syst. Man Cybern.

Part B 37(5), 1212–1225 (2007)

3. Ratha, N.K., Connell, J.H., Bolle, R.M.: Image mosaicing for

rolled ﬁngerprint construction. In: Proceedings of International

Conference on Pattern Recognition, pp. 1651–1653 (1998)

4. Jain, A., Ross, A.: Fingerprint mosaicking. In: Proceedings of

International Conference on Acoustics, Speech, and Signal Pro-

cessing, pp. 4064–4067 (2002)

5. Ross, A., Shah, S., Shah, J.: Image versus feature mosaicing: a

case study in ﬁngerprints. In: Proceedings of SPIE Conference

on Biometric Technology for Human Identiﬁcation III, pp.

620,208-1–620,208-12 (2006)

6. Yang, F., Paindavoine, M., Abdi, H., Monopoly, A.: Development

of a fast panoramic face mosaicing and recognition system. Opt.

Eng. 44(8), 087 005/1–087 005/10 (2005)

7. Lu, X., Jain, A.K.: Pose-robust face recognition using geometry

assisted probabilistic modeling. In: Proceedings of International

Conference on Computer Vision and Pattern Recognition, pp.

502–509 (2005)

8. Kong, S., Heo, J., Abidi, B., Paik, J., M., A.: Recent advances in

visual and infrared face recognition - a review. Comput. Vision

Image Understand. 97(1), 103–135 (2005)

9. Bebis, G., Gyaourova, A., Singh, S., Pavlidis, I.: Face recognition

by fusing thermal infrared and visible imagery. Image Vision

Comput. 24(7), 727–742 (2006)

10. Kong, S., Heo, J., Boug horbel, F., Zheng, Y., Abidi, B.,

Koschan, A., Yi, M., M., A.: Multiscale fusion of visible and

thermal IR images for illumination-invariant face recognition.

Int. J. Comput. Vision 71(2), 215–233 (2007)

11. Singh, R., Vatsa, M., Noore, A.: Integrated multilevel image

fusion and match score fusion of visible and infrared face images

for robust face recognition. Pattern Recognit. 41(3), 880–893

(2008)

12. Lu, X., Jain, A.K.: Integrating range and texture information for

3D face recognition. In: Proceedings of Workshop on Applica-

tions of Computer Vision, pp. 156–163 (2005)

13. Jing, X.Y., Yao, Y.F., Zhang, D., Yang, J.Y., Li, M.: Face and

palmprint pixel level fusion and Kernel DCV-RBF classiﬁer for

small sample biometric recognition. Pattern Recognit. 40(11),

3209–3224 (2007)

14. Noore, A., Singh, R., Vatsa, M.: Robust memor y efﬁcient data

level information fusion of multi-modal biometric images. Inf.

Fusion 8(4), 337–346 (2007)

15. Vatsa, M., Singh, R., Noore, A.: Uniﬁcation of evidence theoretic

fusion algorithms: A case study in level-2 and level-3 ﬁngerprint

features. In: Proceedings of IEEE International Conference on

Biometrics: Theory, Applications, and Systems, pp. 1–6 (2007)

Fusion, User-Specific

JULIAN FIERREZ,JAV IE R ORT EG A-GARCIA

Biometric Recognition Group – ATVS, Escuela

Politecnica Superior, Universidad Autonoma de

Madrid, Campus de Cantoblanco, Madrid, Spain

Synonyms

Adapted fusion; Local fusion; Target-dependent

fusion; User-dependent fusion

Definition

User-speciﬁc fusion in the framework of biometrics,

initially devised for score fusion in the veriﬁcation

mode, refers to techniques used for information fusion

in which there is a speciﬁc fusion function for each

user enrolled in the system. These fusion functions are

retrieved and used fo r information integration in the

same way the enrolled templates corresponding to the

claimed identities are retrieved and used for matching.

User-speciﬁc fusion techniques ﬁnd application in

several biometric fusion scenarios, e.g., multi-modal

fusion, where some subjects may be not adequate for

recognition based on speciﬁc modalities (these evi-

dences can be ignored or given less importance in the

information fusion step), or multi-algorithm fusion,

where some subjects may be better recognized based on

particular algorithms (their fusion functions can be

adapted to give more importance to those algorithms).

The biggest challenge for effective user-speciﬁc

fusion is the need for user-speciﬁc training data,

which is usually very scarce. Recent user-speciﬁc fusion

techniques exploit the usually scarce training data by

considering also for training the information provided

by background users. These new techniques are known

as adapted user-speciﬁc fusion.

System Model

The following nomenclature is used throughout the

essay. Given a multi-biometric veriﬁcation system con-

sisting of a number of uni-modal systems, each one

computes a similarity score between an input bio met-

ric pattern and the enrolled pattern or model of the

Fusion, User-Specific

621

given claimant. The similarity scores are then normal-

ized to a given score range. Let the normalized similar-

ity scores provided by the different uni-modal systems

be combined into a multi-modal score. The design of a

fusion scheme consists in the deﬁnition of a function

which maps a multi-modal score to a fused real value,

so as to maximize the separability of client and impos-

tor fused score distributions. This function may be

ﬁxed or trained (see the entry in this encyclopedia on

Multi-biometrics) by using a set of training scores

(scores known to be genuine or impostor).

The aim in user-speciﬁc fusion is to obtain the best

score fusion function for a particular user, resulting in

the system model shown in Fig. 1.

User-Specific Multi-Biometrics

User-speciﬁc multi-biometric veriﬁcation can be

achieved not only by making the fusion functions

user-speciﬁc as shown in Fig. 1, but also other proces-

sing modules, such as the score normalization and the

decision processing blocks. In the ﬁrst case, each indi-

vidual system will be used as indicated in Fig. 2, in the

latter case the overall system diagram will be as indi-

cated in Fig. 3.

On one hand, user-speciﬁc score normalization has

been traditionally studied for individual behavioral bio-

metric modalities in which there are large variations

between users (such as speech [1] or signature [2]),

Fusion, User-Speciﬁc. Figure 1 System model of multi-biometric verification with user-specific score fusion.

Fusion, User-Speciﬁc. Figure 2 System model of biometric verification with user-specific score normalization.

Fusion, User-Speciﬁc. Figure 3 System model of multi-biometric verification with user-specific decision functions.

622

Fusion, User-Specific

where their application is very effective to compensate

the problems related to the heterogeneity between

users. When user-speciﬁc score normalization is

used in one of the systems being combined in a

multi-biometric setup, the resulting approach can be

seen as integrati ng the multi-biometric data in a user-

speciﬁc way [3]. Despite the success of user-speciﬁc

score normalization in individual modalities, and the

success of fusion techniques, few efforts have been

reported in the literature studying the combined use

of both techniques to make the most out of the usually

scarce user-speciﬁc training data.

On the other hand, the use of user-speciﬁc decisions

in multi-biometrics has been typically studied in

combination with user-speciﬁc score fusion. In this

case, it has been demonstrated that it is better to use

the available training data for computing user-speciﬁc

fusion functions instead of user-speciﬁc decision

schemes [4].

User-Specific Fusion

The idea of exploiting user-speciﬁc parameters at the

score level in multi-modal biometrics was introduced,

to the best of our knowledge, by [5]. In that work, user-

independent weighted linear combination of similarity

scores was demonstrated to be improved by using

either user-speciﬁc weights or user-speciﬁc decision

thresholds, both computed by exhaustive search on

the testing data. The idea of user-speciﬁc fusion para-

meters was also explored by [6]. Other attempts to

personalize multi-modal biometrics include the use

of the claimed identity index as a feature for a global

trained fusion scheme based on neural networks [7],

computing user-speciﬁc weights using lambness

metrics [8], and using personalized Fisher ratios [9].

The existing score fusion approaches can be classi-

ﬁed as global or local depending ﬁrst on the fusion

function (i.e., user-independent or user-speciﬁc fusion

strategies) and secondly on the decision making process

(i.e., user-independent or user-speciﬁc decision thresh-

olds), resulting in [10]: global-learning-global-decision

(GG), local-learning-global-decision (LG), and simi-

larly GL and LL. Some example works on user-speciﬁc

multi-biometrics using this classiﬁcation are: LG [4 , 5,

6, 7, 8, 10, 11], GL [4, 5, 10], and LL [4, 10].

User-speciﬁc score fusion is confronted with a great

challenge: the scarcity of user-speciﬁc training scores.

For overcoming this challenge, the simultaneous use of

user-speciﬁc and background information has been

proposed for training the user-speciﬁc fusion func-

tions, in what has been called adapted user-speciﬁc

fusion. This approach can be seen as a particular case

of a more general type of approaches, referred to as

▶ adapted fusion [11]. In these approaches, a baseline

fusion function is ﬁrst constructed based on some

general knowledge of the problem at hand, and then

adjusted during the operation of the system . The ad-

aptation can be based on ancillary information such as:

the user being claimed (adapted user-speciﬁc fusion),

quality measures of the input biometrics (quality-

based fusion [12], see related entry in this encyclope-

dia), or other kind of environmental information af-

fecting the variou s information channels being fus ed.

Adapted User-Specific Fusion

Adapted methods in the context of user-speciﬁc fusion

refer to the use of both global and local information for

learning the fusion func tions.

The idea of adapted learning is based on the fact

that the amount of available training data in localized

learning is usually not sufﬁcient and representative

enough to guarantee good parameter estimation and

generalization capabilities. To cope with this lack of

robustness derived from partial knowledge of the

problem, one can exploit the information provided

by background global data. In general, the relative

balance between the background information (pool

of users) and the local data (speciﬁc user) is performed

as a tradeoff between both kinds of information.

The system model of adapted user-speciﬁc score

fusion is shown in Fig. 4, where we can see that the

fusion function of a given user is trained with two sets

of training data, both including both genuine and

impostor matc hing scores. The ﬁrst training set con-

sists of scores corresponding to the user being claimed.

The second set consists of scores corresponding to a

pool of background users different to the user being

claimed. By considering these two sets simultane ously,

the resulting adapted user-speciﬁc fusion schemes out-

perform the traditional user-independent fusion (also

known as

▶ global fusion, in which only the pool

of users is used for training), and the traditiona l

user-speciﬁc fusion depicted in Fig. 1 (also known as

▶ local fusion, in which only data from the claimed

Fusion, User-Specific

623

user is used for training). This afﬁrmation has been

demonstrated experimentally in various scenarios,

such as multi-algorithm speaker veriﬁcation [3, 13],

and multi-modal veriﬁcation combining on-line sig-

nature and ﬁngerprint traits [4, 11].

Related Entries

▶ Fusion, Qualit y-Based

▶ Multi-Algorithm Systems

▶ MultiBiometrics

▶ Multi-Modal Systems

References

1. Doddington, G., Liggett, W., Martin, A., Przybocki, M., Reynolds,

D.: Sheeps, goats, lambs and wolves: a statistical analysis of speaker

performance in the NIST 1998 speaker recognition evaluation. In:

Proceedings of International Conference on Speech and Language

Processing, ICSLP (1998)

2. Fierrez-Aguilar, J., Ortega-Garcia, J., Gonzalez-Rodriguez, J.:

Target dependent score normalization techniques and their

application to signature veriﬁcation. IEEE Trans. Syst. Man

Cybernet., Part C 35 , 418–425 (2005)

3. Poh, N., Kittler, J.: Incorporating model-speciﬁc score distribu-

tion in speaker veriﬁcation systems. IEEE Trans. Audio Speech

Lang. Process. 16, 594–606 (2008)

4. Fierrez-Aguilar, J., Garcia-Romero, D., Ortega-Garcia, J., Gonzalez-

Rodriguez, J.: Adapted user-dependent multimodal biometric

authentication exploiting general information. Pattern Recognit.

Lett. 26, 2628–2639 (2005)

5. Jain, A.K., Ross, A.: Learning user-speciﬁc parameters in a multi-

biometric system. In: Proceedings of IEEE International Confer-

ence on Image Processing, ICIP, vol. 1, pp. 57–60 (2002)

6. Wang, Y., Wang, Y., Tan, T.: Combining ﬁngerprint and voice

biometrics for identity veriﬁcation: An experimental comparison.

In: Zhang, D., Jain, A.K. (eds.) Proceedings of International

Conference on Biometric Authentication, ICBA, Springer LNCS-

3072, pp. 663–670 (2004)

7. Kumar, A., Zhang, D.: Integrating palmprint with face for user

authentication. In: Proceedings of Workshop on Multimodal

User Authentication, MMUA, pp. 107–112 (2003)

8. Snelick, R., Uludag, U., Mink, A., Indovina, M., Jain, A.K.: Large

scale evaluation of multimodal biometric authentication using

state-of-the-art systems. IEEE Trans. Pattern Anal. Mach. Intell.

27, 450–455 (2005)

9. Poh, N., Bengio, S.: An investigation of f-ratio client-dependent

normalisation on biometric authentication tasks. In: Proceed-

ings of the IEEE International Conference on Acoustics, Speech

and Signal Processing, ICASSP, vol. 1, pp. 721–724 (2005)

10. Toh, K.A., Jiang, X., Yau, W.Y.: Exploiting local and global

decisions for multimodal biometrics veriﬁcation. IEEE Trans.

Signal Process. 52, 3059–3072 (2004)

11. Fierrez-Aguilar, J., Garcia-Romero, D., Ortega-Garcia, J., Gon-

zalez-Rodriguez, J.: Bayesian adaptation for user-dependent

multimodal biometric authentication. Pattern Recognit. 38,

1317–1319 (2005)

12. Fierrez-Aguilar, J., Ortega-Garcia, J., Gonzalez-Rodriguez, J.,

Bigun, J.: Discriminative multimodal biometric authentication

based on quality measures. Pattern Recognit. 38, 777–779 (2005)

13. Fierrez-Aguilar, J., Garcia-Romero, D ., Ortega-Garcia, J., Gonzalez-

Rodrigez, J.: Speaker veriﬁcation using adapted user-dependent

multilevelfusion.In:Oza,N.C.,Polikar,R.,Kittler,J.,Roli,F.

(eds.) Proceedings of International Workshop on Multiple

Classiﬁer Systems, MCS, Springer LNCS-3541, pp. 356–365 (2005)

Fusion, Wavelet-Based

Wavelet-based fusio n has been widely used in litera-

ture. The wavelet transform is a data analysis tool that

provides a multi-resolution decomposition of an

image. Wavelet-based pixel-level data fusion is used

on two or more sets of probe images. Given two

Fusion, User-peciﬁc. Figure 4 System model of multi-modal biometric authentication with adapted user-specific

score fusion.

624

Fusion, Wavelet-Based

registered images I

and I

of the same object from two

sets of probe images (two different spectral bands in

this case), a two-dimensional discrete wavelet decom-

position is performed on I

and I

to obtain the wavelet

approximation coefﬁcients (a

, a

) and detail coefﬁ-

cients (d

, d

). The wavelet approximation and detail

coefﬁcients of the fused image, a

and d

, are then

calculated as follows:

¼W

 a

þ W

 a

and

¼W

 d

þ W

 d

;

where W

; W

, and W

are weights deter-

mined either empirically or according to some selected

rule. The two-dimensional discrete wavelet inverse

transform is then performed to obtain the fused image.

▶ Multispectral and Hyperspectral Biometrics

Fuzzy Extractor

▶ Encryption, Biometric

▶ Fingerprints Hashing

Fuzzy Vault

Fuzzy vault is where a secret key is hidden behind some

biometric data which are fuzzy and noisy by nature.

▶ Fingerprints Hashing

Fuzzy Vault

625

Gabor Jets

Gabor jets are a set of ﬁlters that are used to extract the

local frequency information from the face images.

These ﬁlters are generally linear ﬁlter with impulse

responses deﬁned by a harmonic function and a

Gaussian function. The Fourier transform of a Gabor

ﬁlter’s impulse response is the convolution of the

Fourier transform of the harmonic function and

the Fourier transform of the Gaussian function.

▶ Face Recognition, Component-Based

Gabor Transform

A complete representation of a signal or image in terms

of coefﬁcients on Gabor wavelets, such that the original

data can be reconstructed exactly by combining togeth-

er those wavelets using their computed coefﬁcients. A

complication is that the necessary coefﬁcients cannot

be obtained simply by operations of ﬁltering or by the

inner product projections of the data with the wavelets,

since they do not constitute an or thogonal basis. More

complex methods are required (biorthogonal bases;

relaxation networks) to obtain the needed expansion

coefﬁcients from projection coefﬁcients. Once

obtained, a Gabor Transform is a powerful tool for

signal or image encoding, analysis, and compression.

▶ Iris Encoding and Recognition using Gabor

Wavelets

Gabor Wavelets

Complex exponentials (Fourier components) multi-

plied by Gaussian envelopes. Although they fail to

satisfy some parts of the stricter mathematical deﬁni-

tions of wavelets, such as orthogonality and compact

support, these elementary functions can constitute a

powerful basis for signal or image encoding, represen-

tation, compression, and analysis. They are increas-

ingly used today in computer vision and in pattern

recognition, particularly in biometrics, where they

are the basis of iris recognition and have also been

used for several other biometric modalities. Among

their advantages (besides forming a complete basis

for signal or image encoding) are; their optimality

under the Heisenberg Uncertainty Principle for simul-

taneous resolution in time/space and in frequency;

their closed analytical form; their self-Fourier property

and closure under convolution and multiplication;

and their neurobiological basis in the receptive ﬁeld

proﬁles of neurons in the mammalian visual cortex.

Their chief disadvantage is that they are not mutually

orthogonal, and so the projection coefﬁcients obtained

by computing their inner product with an image are

not the same as the expansion coefﬁcients that would

be needed to reconstruct the same image exactly

from them.

▶ Face Recognition, Component Based

▶ Iris Encoding and Recognition using Gabo r

Wavelets

▶ Local Image Features

▶ Local Image Filters

2009 Springer Science+Business Media, LLC

Gait

The manner of a person’s movement, speciﬁcally dur-

ing walking is called gait. The human gait cycle consists

of two main phases: during stance phase, the foot is on

the ground, and during the swing phase, the leg is

swinging forward in preparation for the next ground

contact.

▶ Gait, Forensic Evidence of

Gait Analysis

▶ Gait, Forensic Evidence of

Gait Biometrics, Overview

RAMA CHELLAPPA,ASHOK VEERARAGHAVAN

NARAYANAN RAMANATHAN

University of Maryland, College Park, MD, USA

Synonym

Gait recognition

Definition

Gait is deﬁned as the style or manner of walking. Studies

in physchophysics suggest that people can identify famil-

iar individuals using just their gait. This has led to a

number of automated vision based algorithms that use

gait as a biometric. Such a system usually consists of a

video camera capturing images of a person walking with-

in its ﬁeld of view. Appropriate features such as joint

angles or silhouettes are extracted from this video and are

then used to compare with the stored gait signatures of

known individuals. As with any other biometric system,

the system can operate in both the identiﬁcation and

the veriﬁcation mode. Gait as a biometric has several

advantages compared to traditional biometrics such as

ﬁngerprint in that gait is non-intrusive, does not require

cooperation from the individual, and can function at

moderate distances from the subject.

Introduction

The study of human gait has gathered pace in recent

years driven primarily by its potential as a biometric.

Gait-based person authentication has several signiﬁ-

cant advantages compared to traditional biometrics

such as ﬁngerprint or iris. Firstly, gait based biometric

systems do not require the individuals to be cooperative

since the input of these systems is the video feed cap-

tured by passive cameras. Secondly, gait is a non-intru-

sive biometric – it does not require the individuals to

wear any special equipment in order to be recognized.

Thirdly, gait based biometric systems have an extended

range compared to traditional biometrics – they can

operate reliably even when the subjects are tens of

meters away from the camera. Finally, such a system

harnesses the potential of thousands of surveillance

video cameras installed in public locations into a bio-

metric authentication system.

Operation of a Gait Based Biometric

System

The sensor for a gait-based biometric system is a video

camera capturing videos of human subjects walking

within its ﬁeld-of view. The raw sensor video is then

processed to extract relevant features which can then be

used for recognition. If the acquisition conditions are

expected to be controlled and favorable, then the quality

of the video will enable the extraction of features such as

joint angles from the individual video frames. In more

typical uncontrolled settings, the features extracted could

either be background subtracted binary images, silhou-

ettes, shapes or width vectors – all examples of features

capturing the extent of the human body to differing

amounts of detail. During the training phase, several

such sequences of each individual in the gallery ar e col-

lected and the appropriate features are then stored in the

database. During the test phase, each test sequence is

compared with the training sequences available in the

database and the similarity is used to perform person

authentication.

628

Gait

Challenges for Gait Based Biometrics

Systems

The discriminative information in gait is present in

both the shape of the individual and also in the manner

of his/her gait. This means that gait based biometric

systems must be able to model gait as a time series of

features or as a dynamical model in order to perform

accurate recognition. Static template based methods

which have been used for most other biometric systems

need to be adapted to a temporal sequence in order to

achieve robust performance. In this regard, another

challenge is time alignment of two sequences so that

critical events during gait like ‘‘mid-stance’’, ‘‘toe-off ’’

etc. are time aligned accurately so that recognition per-

formance is not affected by inaccurate time alignment

between postures that occur during gait. Since, gait

based person identiﬁcation often occurs without any

particular viewpoint, view-invariance of the feature

extracted from the video is another important challenge.

This will ensure that recognition performance is robust

to changes in the viewpoint of the camera. In scenarios

with moderate amounts of acquisition control, one can

set up multiple video cameras so as to ensure that the

best possible viewpoint which happens to be the fronto-

parallel gait is captured on atleast one of the cameras.

Another challenge for automated gait-based biometrics

is that of changing illumination conditions in the scene.

In order to be robust to changing illumination condi-

tions, background subtraction is typically performed on

the raw videos before the video data is used in a recogni-

tion algorithm. Finally, another important challenge is

the variability in the clothing, shoe type and the surface

on which the individuals walk. Obviously, the clothing of

the subject especially their type of footwear has signiﬁ-

cant impact on the gait features observed and it is im-

portant to bear this in mind while developing gait-based

biometric systems.

Features for Gait Based Biometrics

Silhouette: In most gait-based biometric systems the

cameras can be assumed to be static during the short

duration of time that they capture the gait of a single

individual for veriﬁcation. This allows simple back-

ground models to be built for each of these cameras.

Background subtra ction then identiﬁes the set of

all pi xels in the image that belong to the moving

individual. Figure 1 shows a sequence of color images

captured by a video camera as a person walks through

its ﬁeld of view. Shown below are the binary back-

ground subtracted images in which all pixels belonging

to the individual are white, while the background is

black. This binary image is then scaled to a uniform

size so that the feature extracted is independent of the

distance of the camera from the subject. Several algo-

rithms for gait based person identiﬁcation use this

binary silhouette as a feature [1–5].

Shape: ‘‘Shape is all the geometric information

that remains when location, scale, and rotational

effects are ﬁltered out from the object’’[6]. Kendall’s

statistical shap e is a sparse descriptor of the shape that

describes the shape conﬁguration of k landmark points

in an m-dimensional space as a km matrix contain-

ing the coordinates of the landmarks. Image space is

2-dimensional and therefore it is convenient to de-

scribe the shape vector as a k dimensional complex

vector. First, a binarized silhouette denoting the extent

of the object in an image is obtained. A shape feature

is then extracted from this binarized silhouette. This

feature vector must be invariant to translation and

scaling since the object’s identity should not depend

on the distance of the object from the camera. So any

feature vector that we obtain must be invariant to

translation and scale. This yields the pre-shape of the

object in each frame. Pre-shape is the geometric infor-

mation that remains when location and scale effects are

ﬁltered out. Let the conﬁguration of a set of k landmark

points be given by a k-dimensional complex vector

containing the positions of landmarks. Let us denote

this conﬁguration as X. Centered pre-shape is obtained

by subtracting the mean from the conﬁguration and

then scaling to norm one. The centered pre-shape is

given by

k CX k

; where C ¼ I



; ð1Þ

where I

is a k  k identity matrix and 1

is a

k dimensional vector of ones.

The advantage of using shape feature is that the

differential geometric properties of the spherical man-

ifold in which the shapes lie are very well understood

and therefore, appropriate distance measures that can

account for translational, rotational and scale invar-

iances are well deﬁned. For example, consider two

complex conﬁgurations X and Y with corresponding

preshapes a and b. The full Procrustes distance

Gait Biometrics, Overview

629