Research Article Average Gait Differential Image Based Human...

Research ArticleAverage Gait Differential Image Based Human Recognition

Jinyan Chen1 and Jiansheng Liu2

1 School of Computer Software Tianjin University Tianjin 300072 China2 College of Science Jiangxi University of Science and Technology Ganzhou 330200 China

Correspondence should be addressed to Jinyan Chen chenjinyantjueducn

Received 4 March 2014 Accepted 25 March 2014 Published 6 May 2014

Academic Editor Fei Yu

Copyright copy 2014 J Chen and J Liu This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The difference between adjacent frames of human walking contains useful information for human gait identification Based onthe previous idea a silhouettes difference based human gait recognition method named as average gait differential image (AGDI)is proposed in this paper The AGDI is generated by the accumulation of the silhouettes difference between adjacent frames Theadvantage of thismethod lies in that as a feature image it can preserve both the kinetic and static information of walking Comparingto gait energy image (GEI) AGDI is more fit to representation the variation of silhouettes during walking Two-dimensionalprincipal component analysis (2DPCA) is used to extract features from the AGDI Experiments on CASIA dataset show that AGDIhas better identification and verification performance than GEI Comparing to PCA 2DPCA is a more efficient and less memorystorage consumption feature extraction method in gait based recognition

1 Introduction

With the development of information and Internet tech-nology it is very necessary to authenticate and authorizehuman securely The rapid growth of e-commerce also needsa reliable identification method to ensure safety transactionAs a promising authentication method biometrics is attract-ing more and more attention Biometrics overcomes theinherent flaws and limitations of conventional identificationtechnology and brings a highly secure identification andauthentication method Traditional biometrical resourcesinclude fingerprint face and iris which have been widelyused for authentication However these biometrical featureshave the following disadvantage (1) These features cannotbe taken in a relative long distance (2) Userrsquos cooperation isrequired to get good results As a new biometricsmethod gaitbased human identification overcomes the above limitationand is attracting more and more researchers

Human gait is the manner of one walking which wasfirstly studied in medical field Doctors analyzed the humangait to find out whether patients had health problems [1 2]Later researchers found that just like fingerprint and irisalmost everyone has his distinctive walking style [3 4] So itwas believed that gait could also be used as a biological feature

to recognize the person Although suffering from clothingshoes view angel or environmental context human gait isstill a promising identification method

Humanwalking can be considered as an images sequencehowevermost of the currentmodel-free gait based identifica-tion methods extract features from image sequence withoutconsidering its contained spatiotemporal information Themethod proposed in this paper focuses on the differenceamong the images sequence while constructing the featureimage The procedure can be described as follows Thesilhouettes were normalized to the same height and alignedby the centroid Then the difference between two adjacentsilhouettes was accumulated to get the average gait differen-tial image (AGDI) which is used as the feature image of onewalking Two-dimensional principal component analysis isused to extract feature from AGDI

2 Related Work and Our Contribution

Usually recognition based on human gait includes severaldifferent approaches like walking running and jumping Inthis paperwewould like to restrict the recognition towalkingCurrently human gait recognition can be divided into twocategories model-based methods and motion-based ones

Hindawi Publishing Corporatione Scientific World JournalVolume 2014 Article ID 262398 8 pageshttpdxdoiorg1011552014262398

2 The Scientific World Journal

Model-based approaches aim to describe human move-ment using a mathematical model Cunado et al [5] usedHough transform to extract the positions of arms legs andtorso and then use articulated pendulum to match thosemoving body parts Yoo et al [6] divided the body intohead neck waist leg and arm by image segmentation andthen got the moving curves of these body parts respectivelyLee and Grimson [7] used 7 ellipses to model the humanbody and applied the ellipsesrsquo movement features to identifyhuman Yam et al [8 9] used dynamically coupled oscillatorto describe and analyze the walking and running style ofa person Tafazzoli and Safabakhsh [10] constructed move-ments model based on anatomical proportions then Fouriertransform was used to analyze human walking

Model-freemethods focused on the statistics informationderived from the human gait Cheng et al [11] took HMMand manifold to analyze the relationship between the humanand their gait images Chen et al [12] used parallel HMMto describe the features of human gait Kale et al [13] usedldquofriezerdquo patterns to get features from image sequence and usethem to identify a human Murase and Sakai [14] speeded upthe comparison of human gait by parametric eigenspace rep-resentation Little and Boyd [15] derived scale-independentscalar features from optical flow information of walkingfigures to recognize individuals Wang et al [16] extractedfeature by unwrapping the outer contour of silhouette and usePCA to reduce the dimension of the feature Lee et al [17]adopted product of Fourier coefficients as a distance measurebetween contours to recognize gait Hu [18] combined theenhanced Gabor (EG) representation of the gait energyimage and the regularized local tensor discriminate analysis(RLTDA) method in human identification Hong et al [19]proposed probabilistic framework to identify a humanWanget al [20] proposed spatiotemporal information analysisto get the features of human walking Collins et al [21]extract key frames from the image sequence and comparethe key frames similarity by normalized correlation Sarkaret al [22] estimated the similarity between the galleryimage sequence and the probe image sequence by directlycomputing the correlation between the frame pairs Chen[23] proposes image correlation based human identificationmethod

Our method is similar to gait energy image (GEI) pro-posed by Yu et al [24] Han and Bhanu [25] and framedifference energy image (FDEI) proposed by Chen et al[26] The major difference lies in the approach to generatethe feature image GEI is obtained by directly adding upevery normalized silhouette FDEI is obtained by taking thedifference from every adjacent two frames and then com-bined with the ldquodenoisedrdquo GEI In this paper the differencebetween every two frames will be accumulated to generateaverage gait differential image We also enhanced the featureextraction method by using 2DPCA which has been used inthe application of face recognition [27]

In comparison with the works of state of the art thecontributions of this paper are as follows

Gait Representation Method We propose a new gait featurerepresentation which is called average gait differential image

Centroid

Figure 1 The centroid of a silhouette

Comparing to GEI our method has the advantage of betterperformance

Feature Extraction Method Two-dimensional principal com-ponent analysis (2DPCA) is used to extract features fromAGDI which can be more efficient and save more storagecomparing to the widely used one-dimensional principalcomponent analysis (PCA)

3 Average Gait Differential Image(AGDI) Representation

31 The Construction of AGDI The construction of averagedifferential image can be expressed in the following steps

Silhouette Segmentation Gauss model is used to get thebackground model from the original images sequence Toeliminate the effect of noise every image is blurred by Gaussfilter The method proposed by Wang et al [16] is used toextract walking object from the original images

NormalizationTo exclude the distance effect every silhouetteis normalized to the same height using bicubic interpolation

Alignment and SubtractionWe define the centroid (119909119888 119910119888) of

a silhouette as follows

119909119888=1

119899

119899

sum119894=1

119909119894

119910119888=1

119899

119899

sum119894=1

119910119894

(1)

where 119899 is the number of pixels in the silhouette Figure 1shows the centroid of a silhouette

The Scientific World Journal 3

Figure 2 Differential images and average gait differential image

Suppose that 119868119895and 119868119895+1

are two adjacent images alignedby the centroid the gait differential image 119863

119895can then be

defined as follows

119863119895(119909 119910) =

0 if 119868119895(119909 119910) = 119868

119895+1(119909 119910)

1 if 119868119895(119909 119910) = 119868

119895+1(119909 119910)

(2)

where 119895 is the frame number in the image sequence and 119909 and119910 are values in the 2D image coordinate

Get the Average Gait Differential Image By overlapping all thedifferential images of one human gait cycle we can get thefollowing average gait differential image

119866 (119909 119910) =1

119873 minus 1

119873minus1

sum119895=1

119863119895(119909 119910) (3)

where119873 is the number of frames in the complete gait cycle(s)of a silhouette sequence Figure 2 show the differential imagesin a gait cycle and the average gait differential imagerespectively

32 Feature Extraction Although intheprevious section wehave compressed the human gait features into one imagethe dimensionality of the average gait differential imageis still very large The most commonly used dimensionalreduction method is principal component analysis (PCA)In traditional PCA method every two-dimensional imagemust be transformed into one-dimensional vector leadingto a covariance matrix with large size This large matrix willuse massive memory storage and is difficult to be evaluatedaccurately

To reduce memory storage and speed up the calculationthis paper adopts the two-dimensional principal componentanalysis (2DPCA) to reduce the dimensionality which wasfirst proposed by Yang et al [27] in the recombination ofhuman face Our final target is to project average gait differ-ential image 119866 a119898 times 119899 random matrix onto a119898-dimensionprojected vector119884which is called the projected feature vectorof image 119866 by the following linear transformation [27]

119884 = 119866119882 (4)

where 119882 denotes a 119899-dimensional unitary column vectorTo preserve the features of 119866 119882 should make 119884 have the

maximum scatter We define 119878119910as the scatter of 119884 [27] as

follows

119878119910= 119864 (119884 minus 119864 (119884)) (119884 minus 119864 (119884))

119879

= 119864 (119866119882 minus 119864 (119866119882)) (119866119882 minus 119864 (119866119882))119879

= 119864 (119866 minus 119864 (119866))119882119882119879

(119866 minus 119864 (119866))119879

(5)

The trace of 119878119910can be expressed as

tr (119878119910) = 119882

119879

(119864(119866 minus 119864 (119866))119879

(119866 minus 119864 (119866)))119882 (6)

Here we can define the image covariance matrix as

119862119905= 119864(119866 minus 119864 (119866))

119879

(119866 minus 119864 (119866)) (7)

In this paper average gait differential image for eachindividual (1 2 119872) is expressed as 119866

1 1198662sdot sdot sdot 119866119872 and

then 119862119905can be calculated by

119862119905=

119872

sum119894=1

(119866119894minus 119866)119879

(119866119894minus 119866) (8)

Our target is to find a series of 119882opt in formula (6) tomake tr (119878

119910) have the maximum value According to [13]

the optimal projection axis 119882opt is the unitary orthogonaleigenvector of 119862

119905corresponding to the largest eigenvalue

We define the first 119889 unitary orthogonal eigenvector as11988211198822 119882

119889 that is

1198821 119882

119889 = arg max (119882119879119862

119905119882)

119882119894119882119895= 0 119894 = 119895 119894 119895 = 1 119889

119882119894119882119895= 1 119894 = 119895 119895 = 1 119889

(9)

The first d optimal projection vectors 1198821 119882

119889 are

used to extract features from the average different imagesThat is to say given an average gait differential image 119883 let

119884119896= 119866119882

119896 119896 = 1 2 119889 (10)

Then we get a series of projected feature vectors1198841sdot sdot sdot 119884119889

which are different from those scalar counterparts obtainedfrom PCA By using 2DPCA the original 119898 times 119899 image isprojected to a119898 times 119889 (119889 le 119899) feature matrix 119884 as

119884 =[[

[

1198841

119884119889

]]

]

(11)


k = 1 k = 2 k = 3 k = 5 k = 10 k = 20 Original images

Figure 3 The reconstructed subimages (119896 = 1 2 3 5 10 20) and the original images

The distance between two feature matrixes is defined as

119889 (119884 (119894) 119884 (119895)) =

119889

sum119896=1

100381710038171003817100381710038171003817119884(119894)

119896minus 119884(119895)

119896

100381710038171003817100381710038171003817 (12)

where 119884(119894)119896minus119884(119895)

119896means the Euclidean distance between two

vectors

33 Identification and Verification Following the patternproposed by Sarkar et al [22] we evaluate performance forboth identification and verification scenarios

In the scenario of identification every images sequencein the gallery (training set) is transformed to a119898 times 119889 featurematrix 119884(119894) by the method described in Section 32 Given aprobe silhouette sequence its transformed feature matrix isdefined as 119875 This probe 119875 is assigned to person 119896 by usingthe nearest neighbor method

119889 (119884 (119896) 119875) = min119894

119889 (119884 (119894) minus 119875) (13)

In the scenario of verification the similarity between twofeature matrixes is defined as the negative of distance that is

Sim (119884 (119894) 119884 (119895)) = minus119889

sum119896=1

100381710038171003817100381710038171003817119884(119894)

119896minus 119884(119895)

119896

100381710038171003817100381710038171003817 (14)

In this paper the similarity between a probe 119875119895 and 119884(119894)

in the gallery is defined as 119911-normed similarity [28]

Sim (119875119895 119884 (119894)) =

Sim (119875119895 119884 (119894)) minusMean

119894Sim (119875

119895 119884 (119894))

sd119894Sim (119875

119895 119884 (119894))

(15)

where sd is standard deviationFAR (false acceptance rate) FRR (false rejection rate) and

EER (equal error rate) are used to evaluate the performanceof verification [22]

34 DPCA-Based Average Gait Differential Image Reconstruc-tion In PCA the principal components and eigenvectors canbe combined to reconstruct the original matrix Similarly2DPCA can also be used to reconstruct an average gaitdifferential silhouette

Suppose that the eigenvectors corresponding to thelargest 119889 eigenvectors of 119862

119905are119882

1 119882

119889 that is

119884119896= 119866119882

119896 119896 = 1 119889

[1198841sdot sdot sdot 119884119889] = 119883 [119882

1sdot sdot sdot119882119889]

(16)

According to formula (9)1198821sdot sdot sdot119882119889are normal orthog-

onal vectors so the new reconstructed image 119866 can beexpressed as

119866 = 119866 [1198821sdot sdot sdot119882119889] [1198821sdot sdot sdot119882119889]119879

= [1198841sdot sdot sdot 119884119889] [1198821sdot sdot sdot119882119889]119879

=

119889

sum119896=1

119884119896119882119896

119879

(17)

4 Experiments and Analysis

41 Data and Parameters CASIA gait database (Dataset B)[29] one of the largest gait databases in gait-research fieldis used in the following experiment Dataset B consists of124 subjects (93 males and 31 females) captured from 11view angles (ranging from 0 to 180∘ degree with view angleinterval of 18) For every person there are six normal walkingsequences (named normal-01sdot sdot sdot normal-06) conducted fromevery view angle Every walking sequence contains 3ndash8 gaitcycles (about 40ndash100 frames)Thevideo frame size is 320times240pixels and the frame rate is 25 fps We use all the 124 objectsin Dataset B to carry out our experiments

In all the following experiments 2DPCA method wasused to get features from the images and 20 eigenvectorscorresponding to the first 20 eigenvalues are used to producefeatures (119889 = 20)The size of original image is 240times320 exceptfor special declaration

For each person from every view angle we select the 39frames from sequence normal-01 as the training data (gallery)and 13 frames (except for special declaration) from sequencenormal-02 as the test data (probe)

For every view angle each time we leave one trainingimage sequence out and use the remainder as the trainingset In the scenario of identification we calculate the distance


between the probe corresponding to the leave out trainingimage sequence and the 124 classes (including the leave outimage sequence) In the scenario of verification we calculatethe similarity between the probe corresponding to the leaveout training image sequence and the 124 classes (including theleave out image sequence)

42 The Reconstruction of Subimage Formula (17) indicatesthat we can reconstruct the subimage from the 119882

119896and

119884119896 Figure 3 shows the result of the reconstruction For the

consideration of illustration we normalize the brightness ofevery image into the range of 0ndash255

As showed in Figure 3 the first and the second (119896 = 1119896 = 2 in formula (17)) subimages corresponding to largeeigenvectors of 119862

119905contain the most energy of the original

images With the increase of 119896 the subimage contains moredetailed information

We also demonstrate the eigenvalue calculated by2DPCA Figure 4 shows themagnitude of the eigenvalues thatquickly converges to zero

43 Performance Evaluation

431 Comparison of AGDI and GEI We compare the per-formance of our AGDI base method with that of gait energyimage based method (In this paper we use real templatefor GEI method [25]) Table 1 shows the rank 1 and rank 5identification rates comparing with GEI

To compare the performance of verification we alsoevaluate the FAR (false acceptance rate) and FRR (false rejec-tion rate) for AGDI and GEI The ROC (receiver operatingcharacteristic) curves under view angles 0∘ 90∘ and 180∘ areshown in Figures 5(a)ndash5(c) The comparison of EERs (equalerror rate) is shown in Figure 5(d)

From Table 1 and Figure 5 we can see that almost underevery view angle AGDI has better performance comparing toGEI (except that it is comparable under 0 view angel)

As also can be seen in Table 2 the best performance wasobtained from the walking sequence taken from 0∘ 90∘ and180∘ while the worst was obtained from thewalking sequencetaken from 36∘ and 54∘This is probably due to the least visualdeformation in the former degrees butmore in the latter ones

432 The Effect of Images Amount From the definition ofAGDI (formula (3)) we can see that the AGDI image is theaverage value of differential images It should be expected thatthe use of more images as sample would contribute to a moreprecise result To demonstrate this effect a test was conductedby selectively choosing 13 26 and 39 (approximately corre-sponding to 1 2 and 3 gait cycles) images from 90 degree insequences normal 0l-02 as test dataset probe Figure 5 showsthe experimental result

Indeed in Figure 5 the performance of 26 and 39 imagesis much better than that of 13

433 Comparison of 2DPCA and PCA We also design anexperiment to compare the performance of 2DPCA andPCA which were applied in the step of feature extraction

Table 1 Comparison of identification performance of AGDI andGEI

Viewangel

Rank 1 performance Rank 5 performanceAGDI GEI AGDI GEI

0∘ 72 68 88 8918∘ 54 37 73 6036∘ 35 22 51 4454∘ 44 26 55 4572∘ 66 44 86 6690∘ 81 77 93 90108∘ 78 62 92 85126∘ 46 36 73 53144∘ 49 34 72 52162∘ 59 35 75 48180∘ 88 84 94 93

0 100 200 300 4000

50

100

150

200

250

300

350Ei

genv

alue

s

Number of eigenvalues

Figure 4 The magnitude of eigenvalue

respectively The data set view angel is 90∘ and every frameis resized to 120 times 160

As illustrated in Figure 7 the performance of 2DPCAachieving the maximum at about 25 dimensions is muchbetter than PCA

The key step for both PCA and 2DPCA is to get theeigenvalue and eigenvector from the covariance matrix CtFor PCA method every line of the covariance matrix cor-responds to an image as does the whole covariance matrixfor the 2DPCA That is if the image size is 119898 times 119899 for PCAthe covariance matrix will be an (119898 times 119899) times (119898 times 119899) matrixwhile for 2DPCA it is just an 119898 times 119898 matrix We resize thesilhouette to different sizes and compare the CPU time ofPCA and 2DPCA for the step of feature extraction

From Table 2 we can see that 2DPCA is more efficientthan PCA especially when the image is large


0 02 04 06 08 10

02

04

06

08

1FA

R

FRR

EER

AGDIGEI

(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

AGDIGEI

(b)

AGDIGEI

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

(c)

0 18 36 54 72 90 108 126 144 162 1800

005

01

015

02

025

03EE

R

View angle

AGDIGEI

(d)

Figure 5 (a)ndash(c) The comparison of ROC curves of AGDI and GEI with view angles 0∘ 90∘ and 180∘ respectively (d) The comparison ofEERs of AGDI and GEI with view angles 0∘ndash180∘

Table 2 Comparison of CPU time (ms) for PCA and 2DPCA feature extraction (CPU Intel Core i3 230GHz RAM 4GB)

Feature extraction method Image size32 times 24 64 times 48 96 times 72 128 times 96 160 times 120 192 times 144

2DPCA 17ms 47ms 105ms 167ms 257ms 431msPCA 117ms 318ms 1273ms 4288ms 8896ms 17876ms

5 Conclusions

An average gait differential image based human recognitionmethod is proposed in this paper (Figure 6) The Kernelidea of AGDI is to apply the average of differential imageas the feature image and use the two-dimensional principal

component analysis to extract features Experiments onCASIA dataset show the following (1) Comparing to GEIAGDI method achieves better identification and verificationperformance (2) Comparing to PCA 2DPCA is more effi-cient and needs lower memory storage (3)The 0 90 and 180degrees silhouettes are more fit to AGDI base recognition


2 4 6 8 1075

80

85

90

95

100Re

cogn

ition

accu

racy

()

Rank

13 differential images26 differential images39 differential images

(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER


(b)

Figure 6 (a) Recognition accuracy of different probe sizes (b) The comparison of ROC curves of different probe sizes

5 10 15 20 25 300

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

)

Dimension

2DPCAPCA

Figure 7 The rank 1 recognition accuracy comparison of 2DPCAand PCA

Conflict of Interests

The authors declare that they have no financial or personalrelationships with other people or organizations that caninappropriately influence their work There are no profes-sional or other personal interests of any nature or kind in anyproduct service andor company that could be construed asinfluencing the position presented in or the review of thepaper

Acknowledgments

This work is funded by PhD Programs Foundation ofMinistry of Education of China (no 20100032120011) Datasetused in this paper is provided by Institute of AutomationChinese Academy of Sciences [29]

References

[1] M W Whittle ldquoClinical gait analysis a reviewrdquo Human Move-ment Science vol 15 no 3 pp 369ndash387 1996

[2] S Lobet C Detrembleur F Massaad and C Hermans ldquoThree-dimensional gait analysis can shed new light on walking inpatients with haemophiliardquo The Scientific World Journal vol2013 Article ID 284358 7 pages 2013

[3] S V Stevenage M S Nixon and K Vince ldquoVisual analysis ofgait as a cue to identityrdquo Applied Cognitive Psychology vol 13no 6 pp 513ndash526 1999

[4] C Ben Abdelkader R Cutler and L Davis ldquoStride and cadenceas a biometric in automatic person identification and verifica-tionrdquo in Proceedings of the 5th IEEE International Conference onAutomatic Face and Gesture Recognition (FGR rsquo02) 2002

[5] D CunadoM S Nixon and J N Carter ldquoAutomatic extractionand description of human gait models for recognition pur-posesrdquo Computer Vision and Image Understanding vol 90 no1 pp 1ndash41 2003

[6] J H Yoo D Hwang and M S Nixon ldquoGender classificationin human gait using support vector machinerdquo in AdvancedConcepts for Intelligent Vision Systems 7th International Con-ference ACIVS 2005 Antwerp Belgium September 20-23 2005Proceedings vol 3708 of Lecture Notes in Computer Science pp138ndash145 Springer 2005

[7] L Lee and W E L Grimson ldquoGait analysis for recognitionand classificationrdquo in Proceedings of 5th IEEE International


Conference on Automatic Face Gesture Recognition pp 155ndash1622002

[8] C Y C Yam M S Nixon and J N Carter ldquoAutomatedperson recognition by walking and running via model-basedapproachesrdquo Pattern Recognition vol 37 no 5 pp 1057ndash10722004

[9] C Yam M S Nixon and J N Carter ldquoGait recognition bywalking and running a model-based approachrdquo in Proceedingsof the 5th Asian Conference on Computer Vision MelbourneAustralia 2002

[10] F Tafazzoli andR Safabakhsh ldquoModel-based human gait recog-nition using leg and arm movementsrdquo Engineering Applicationsof Artificial Intelligence vol 23 no 8 pp 1237ndash1246 2010

[11] M-H Cheng M-F Ho and C-L Huang ldquoGait analysis forhuman identification through manifold learning and HMMrdquoPattern Recognition vol 41 no 8 pp 2541ndash2553 2008

[12] C Chen J Liang H Zhao H Hu and J Tian ldquoFactorial HMMand parallel HMM for gait recognitionrdquo IEEE Transactions onSystems Man and Cybernetics C Applications and Reviews vol39 no 1 pp 114ndash123 2009

[13] A Kale A Roy-Chowdhury and R Chellappa ldquoGait-basedhuman identification from a monocular video sequencerdquo inHandbook on Pattern Recognition and Computer Vision C HCheng and P S P Wang Eds World Scientific PublishingCompany 3rd edition 2005

[14] H Murase and R Sakai ldquoMoving object recognition ineigenspace representation gait analysis and lip readingrdquoPatternRecognition Letters vol 17 no 2 pp 155ndash162 1996

[15] J J Little and J E Boyd ldquoRecognizing people by their gait theshape of motionrdquo Videre Journal of Computer Vision Researchvol 1 no 2 1998

[16] LWang T Tan H Ning andWHu ldquoSilhouette analysis-basedgait recognition for human identificationrdquo IEEE Transactionson Pattern Analysis and Machine Intelligence vol 25 no 12 pp1505ndash1518 2003

[17] C P Lee A W C Tan and S C Tan ldquoGait recognition viaoptimally interpolated deformable contoursrdquo Pattern Recogni-tion Letters vol 34 no 6 pp 663ndash669 2013

[18] H Hu ldquoEnhanced gabor feature based classification using aregularized locally tensor discriminant model for multiviewgait recognitionrdquo IEEE Transactions on Circuits and Systems forVideo Technology vol 23 no 7 pp 1274ndash1286 2013

[19] S Hong H Lee and E Kim ldquoProbabilistic gait modelling andrecognitionrdquo Computer Vision vol 7 no 1 pp 56ndash70 2013

[20] C Wang J Zhang L Wang J Pu and X Yuan ldquoHuman iden-tification using temporal information preserving gait templaterdquoIEEE Transactions on Pattern Analysis andMachine Intelligencevol 34 no 11 pp 2164ndash2176 2012

[21] R T Collins R Gross and S Jianbo ldquoSilhouette-based humanidentification from body shape and gaitrdquo in Proceedings of5th IEEE International Conference on Automatic Face GestureRecognition 2002

[22] S Sarkar P J Phillips Z Liu I R Vega P Grother andK W Bowyer ldquoThe humanID gait challenge problem datasets performance and analysisrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 27 no 2 pp 162ndash1772005

[23] J Chen ldquoGait correlation analysis based human identificationrdquoThe Scientific World Journal vol 2014 Article ID 168275 8pages 2014

[24] S Yu T Tan KHuang K Jia andXWu ldquoA study on gait-basedgender classificationrdquo IEEE Transactions on Image Processingvol 18 no 8 pp 1905ndash1910 2009

[25] J Han and B Bhanu ldquoIndividual recognition using gait energyimagerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 28 no 2 pp 316ndash322 2006

[26] C Chen J LiangH ZhaoHHu and J Tian ldquoFrame differenceenergy image for gait recognition with incomplete silhouettesrdquoPattern Recognition Letters vol 30 no 11 pp 977ndash984 2009

[27] J Yang D Zhang A F Frangi and J-Y Yang ldquoTwo-dimensional PCA a new approach to appearance-based facerepresentation and recognitionrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 26 no 1 pp 131ndash1372004

[28] P J Phillips P Grother R Micheals D M Blackburn ETabassi and M Bone ldquoFace recognition vendor test 2002rdquoin Proceedings of the IEEE International Workshop on FaceRecognition 2003

[29] ldquoCASIAGaitDatabaserdquo 2009httpwwwsinobiometricscom

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



Model-based approaches aim to describe human move-ment using a mathematical model Cunado et al [5] usedHough transform to extract the positions of arms legs andtorso and then use articulated pendulum to match thosemoving body parts Yoo et al [6] divided the body intohead neck waist leg and arm by image segmentation andthen got the moving curves of these body parts respectivelyLee and Grimson [7] used 7 ellipses to model the humanbody and applied the ellipsesrsquo movement features to identifyhuman Yam et al [8 9] used dynamically coupled oscillatorto describe and analyze the walking and running style ofa person Tafazzoli and Safabakhsh [10] constructed move-ments model based on anatomical proportions then Fouriertransform was used to analyze human walking

Model-freemethods focused on the statistics informationderived from the human gait Cheng et al [11] took HMMand manifold to analyze the relationship between the humanand their gait images Chen et al [12] used parallel HMMto describe the features of human gait Kale et al [13] usedldquofriezerdquo patterns to get features from image sequence and usethem to identify a human Murase and Sakai [14] speeded upthe comparison of human gait by parametric eigenspace rep-resentation Little and Boyd [15] derived scale-independentscalar features from optical flow information of walkingfigures to recognize individuals Wang et al [16] extractedfeature by unwrapping the outer contour of silhouette and usePCA to reduce the dimension of the feature Lee et al [17]adopted product of Fourier coefficients as a distance measurebetween contours to recognize gait Hu [18] combined theenhanced Gabor (EG) representation of the gait energyimage and the regularized local tensor discriminate analysis(RLTDA) method in human identification Hong et al [19]proposed probabilistic framework to identify a humanWanget al [20] proposed spatiotemporal information analysisto get the features of human walking Collins et al [21]extract key frames from the image sequence and comparethe key frames similarity by normalized correlation Sarkaret al [22] estimated the similarity between the galleryimage sequence and the probe image sequence by directlycomputing the correlation between the frame pairs Chen[23] proposes image correlation based human identificationmethod

Our method is similar to gait energy image (GEI) pro-posed by Yu et al [24] Han and Bhanu [25] and framedifference energy image (FDEI) proposed by Chen et al[26] The major difference lies in the approach to generatethe feature image GEI is obtained by directly adding upevery normalized silhouette FDEI is obtained by taking thedifference from every adjacent two frames and then com-bined with the ldquodenoisedrdquo GEI In this paper the differencebetween every two frames will be accumulated to generateaverage gait differential image We also enhanced the featureextraction method by using 2DPCA which has been used inthe application of face recognition [27]

In comparison with the works of state of the art thecontributions of this paper are as follows

Gait Representation Method We propose a new gait featurerepresentation which is called average gait differential image

Centroid

Figure 1 The centroid of a silhouette

Comparing to GEI our method has the advantage of betterperformance

Feature Extraction Method Two-dimensional principal com-ponent analysis (2DPCA) is used to extract features fromAGDI which can be more efficient and save more storagecomparing to the widely used one-dimensional principalcomponent analysis (PCA)

3 Average Gait Differential Image(AGDI) Representation

31 The Construction of AGDI The construction of averagedifferential image can be expressed in the following steps

Silhouette Segmentation Gauss model is used to get thebackground model from the original images sequence Toeliminate the effect of noise every image is blurred by Gaussfilter The method proposed by Wang et al [16] is used toextract walking object from the original images

NormalizationTo exclude the distance effect every silhouetteis normalized to the same height using bicubic interpolation

Alignment and SubtractionWe define the centroid (119909119888 119910119888) of

a silhouette as follows

119909119888=1

119899

119899

sum119894=1

119909119894

119910119888=1

119899

119899

sum119894=1

119910119894

(1)

where 119899 is the number of pixels in the silhouette Figure 1shows the centroid of a silhouette



Suppose that 119868119895and 119868119895+1


119895can then be

defined as follows

119863119895(119909 119910) =

0 if 119868119895(119909 119910) = 119868

119895+1(119909 119910)

1 if 119868119895(119909 119910) = 119868

119895+1(119909 119910)

(2)



119866 (119909 119910) =1

119873 minus 1

119873minus1

sum119895=1

119863119895(119909 119910) (3)




119884 = 119866119882 (4)



follows

119878119910= 119864 (119884 minus 119864 (119884)) (119884 minus 119864 (119884))

119879

= 119864 (119866119882 minus 119864 (119866119882)) (119866119882 minus 119864 (119866119882))119879

= 119864 (119866 minus 119864 (119866))119882119882119879

(119866 minus 119864 (119866))119879

(5)


tr (119878119910) = 119882

119879

(119864(119866 minus 119864 (119866))119879

(119866 minus 119864 (119866)))119882 (6)


119862119905= 119864(119866 minus 119864 (119866))

119879

(119866 minus 119864 (119866)) (7)




119862119905=

119872

sum119894=1

(119866119894minus 119866)119879

(119866119894minus 119866) (8)






119889 that is

1198821 119882

119889 = arg max (119882119879119862

119905119882)

119882119894119882119895= 0 119894 = 119895 119894 119895 = 1 119889

119882119894119882119895= 1 119894 = 119895 119895 = 1 119889

(9)


119889 are


119884119896= 119866119882

119896 119896 = 1 2 119889 (10)



119884 =[[

[

1198841

119884119889

]]

]

(11)





119889 (119884 (119894) 119884 (119895)) =

119889

sum119896=1

100381710038171003817100381710038171003817119884(119894)

119896minus 119884(119895)

119896

100381710038171003817100381710038171003817 (12)

where 119884(119894)119896minus119884(119895)


vectors



119889 (119884 (119896) 119875) = min119894

119889 (119884 (119894) minus 119875) (13)


Sim (119884 (119894) 119884 (119895)) = minus119889

sum119896=1

100381710038171003817100381710038171003817119884(119894)

119896minus 119884(119895)

119896

100381710038171003817100381710038171003817 (14)



Sim (119875119895 119884 (119894)) =

Sim (119875119895 119884 (119894)) minusMean

119894Sim (119875

119895 119884 (119894))

sd119894Sim (119875

119895 119884 (119894))

(15)





119905are119882

1 119882

119889 that is

119884119896= 119866119882

119896 119896 = 1 119889

[1198841sdot sdot sdot 119884119889] = 119883 [119882


(16)





=

119889

sum119896=1

119884119896119882119896

119879

(17)









119896and
















Viewangel


0∘ 72 68 88 8918∘ 54 37 73 6036∘ 35 22 51 4454∘ 44 26 55 4572∘ 66 44 86 6690∘ 81 77 93 90108∘ 78 62 92 85126∘ 46 36 73 53144∘ 49 34 72 52162∘ 59 35 75 48180∘ 88 84 94 93

0 100 200 300 4000

50

100

150

200

250

300

350Ei

genv

alue

s








0 02 04 06 08 10

02

04

06

08

1FA

R

FRR

EER

AGDIGEI

(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

AGDIGEI

(b)

AGDIGEI

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

(c)

0 18 36 54 72 90 108 126 144 162 1800

005

01

015

02

025

03EE

R

View angle

AGDIGEI

(d)





5 Conclusions




2 4 6 8 1075

80

85

90

95

100Re

cogn

ition

accu

racy

()

Rank


(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER


(b)


5 10 15 20 25 300

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

)

Dimension

2DPCAPCA




Acknowledgments


References


































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











Suppose that 119868119895and 119868119895+1


119895can then be

defined as follows

119863119895(119909 119910) =

0 if 119868119895(119909 119910) = 119868

119895+1(119909 119910)

1 if 119868119895(119909 119910) = 119868

119895+1(119909 119910)

(2)



119866 (119909 119910) =1

119873 minus 1

119873minus1

sum119895=1

119863119895(119909 119910) (3)




119884 = 119866119882 (4)



follows

119878119910= 119864 (119884 minus 119864 (119884)) (119884 minus 119864 (119884))

119879

= 119864 (119866119882 minus 119864 (119866119882)) (119866119882 minus 119864 (119866119882))119879

= 119864 (119866 minus 119864 (119866))119882119882119879

(119866 minus 119864 (119866))119879

(5)


tr (119878119910) = 119882

119879

(119864(119866 minus 119864 (119866))119879

(119866 minus 119864 (119866)))119882 (6)


119862119905= 119864(119866 minus 119864 (119866))

119879

(119866 minus 119864 (119866)) (7)




119862119905=

119872

sum119894=1

(119866119894minus 119866)119879

(119866119894minus 119866) (8)






119889 that is

1198821 119882

119889 = arg max (119882119879119862

119905119882)

119882119894119882119895= 0 119894 = 119895 119894 119895 = 1 119889

119882119894119882119895= 1 119894 = 119895 119895 = 1 119889

(9)


119889 are


119884119896= 119866119882

119896 119896 = 1 2 119889 (10)



119884 =[[

[

1198841

119884119889

]]

]

(11)





119889 (119884 (119894) 119884 (119895)) =

119889

sum119896=1

100381710038171003817100381710038171003817119884(119894)

119896minus 119884(119895)

119896

100381710038171003817100381710038171003817 (12)

where 119884(119894)119896minus119884(119895)


vectors



119889 (119884 (119896) 119875) = min119894

119889 (119884 (119894) minus 119875) (13)


Sim (119884 (119894) 119884 (119895)) = minus119889

sum119896=1

100381710038171003817100381710038171003817119884(119894)

119896minus 119884(119895)

119896

100381710038171003817100381710038171003817 (14)



Sim (119875119895 119884 (119894)) =

Sim (119875119895 119884 (119894)) minusMean

119894Sim (119875

119895 119884 (119894))

sd119894Sim (119875

119895 119884 (119894))

(15)





119905are119882

1 119882

119889 that is

119884119896= 119866119882

119896 119896 = 1 119889

[1198841sdot sdot sdot 119884119889] = 119883 [119882


(16)





=

119889

sum119896=1

119884119896119882119896

119879

(17)









119896and
















Viewangel


0∘ 72 68 88 8918∘ 54 37 73 6036∘ 35 22 51 4454∘ 44 26 55 4572∘ 66 44 86 6690∘ 81 77 93 90108∘ 78 62 92 85126∘ 46 36 73 53144∘ 49 34 72 52162∘ 59 35 75 48180∘ 88 84 94 93

0 100 200 300 4000

50

100

150

200

250

300

350Ei

genv

alue

s








0 02 04 06 08 10

02

04

06

08

1FA

R

FRR

EER

AGDIGEI

(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

AGDIGEI

(b)

AGDIGEI

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

(c)

0 18 36 54 72 90 108 126 144 162 1800

005

01

015

02

025

03EE

R

View angle

AGDIGEI

(d)





5 Conclusions




2 4 6 8 1075

80

85

90

95

100Re

cogn

ition

accu

racy

()

Rank


(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER


(b)


5 10 15 20 25 300

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

)

Dimension

2DPCAPCA




Acknowledgments


References


































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation













119889 (119884 (119894) 119884 (119895)) =

119889

sum119896=1

100381710038171003817100381710038171003817119884(119894)

119896minus 119884(119895)

119896

100381710038171003817100381710038171003817 (12)

where 119884(119894)119896minus119884(119895)


vectors



119889 (119884 (119896) 119875) = min119894

119889 (119884 (119894) minus 119875) (13)


Sim (119884 (119894) 119884 (119895)) = minus119889

sum119896=1

100381710038171003817100381710038171003817119884(119894)

119896minus 119884(119895)

119896

100381710038171003817100381710038171003817 (14)



Sim (119875119895 119884 (119894)) =

Sim (119875119895 119884 (119894)) minusMean

119894Sim (119875

119895 119884 (119894))

sd119894Sim (119875

119895 119884 (119894))

(15)





119905are119882

1 119882

119889 that is

119884119896= 119866119882

119896 119896 = 1 119889

[1198841sdot sdot sdot 119884119889] = 119883 [119882


(16)





=

119889

sum119896=1

119884119896119882119896

119879

(17)









119896and
















Viewangel


0∘ 72 68 88 8918∘ 54 37 73 6036∘ 35 22 51 4454∘ 44 26 55 4572∘ 66 44 86 6690∘ 81 77 93 90108∘ 78 62 92 85126∘ 46 36 73 53144∘ 49 34 72 52162∘ 59 35 75 48180∘ 88 84 94 93

0 100 200 300 4000

50

100

150

200

250

300

350Ei

genv

alue

s








0 02 04 06 08 10

02

04

06

08

1FA

R

FRR

EER

AGDIGEI

(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

AGDIGEI

(b)

AGDIGEI

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

(c)

0 18 36 54 72 90 108 126 144 162 1800

005

01

015

02

025

03EE

R

View angle

AGDIGEI

(d)





5 Conclusions




2 4 6 8 1075

80

85

90

95

100Re

cogn

ition

accu

racy

()

Rank


(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER


(b)


5 10 15 20 25 300

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

)

Dimension

2DPCAPCA




Acknowledgments


References


































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












119896and
















Viewangel


0∘ 72 68 88 8918∘ 54 37 73 6036∘ 35 22 51 4454∘ 44 26 55 4572∘ 66 44 86 6690∘ 81 77 93 90108∘ 78 62 92 85126∘ 46 36 73 53144∘ 49 34 72 52162∘ 59 35 75 48180∘ 88 84 94 93

0 100 200 300 4000

50

100

150

200

250

300

350Ei

genv

alue

s








0 02 04 06 08 10

02

04

06

08

1FA

R

FRR

EER

AGDIGEI

(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

AGDIGEI

(b)

AGDIGEI

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

(c)

0 18 36 54 72 90 108 126 144 162 1800

005

01

015

02

025

03EE

R

View angle

AGDIGEI

(d)





5 Conclusions




2 4 6 8 1075

80

85

90

95

100Re

cogn

ition

accu

racy

()

Rank


(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER


(b)


5 10 15 20 25 300

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

)

Dimension

2DPCAPCA




Acknowledgments


References


































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










0 02 04 06 08 10

02

04

06

08

1FA

R

FRR

EER

AGDIGEI

(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

AGDIGEI

(b)

AGDIGEI

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER

(c)

0 18 36 54 72 90 108 126 144 162 1800

005

01

015

02

025

03EE

R

View angle

AGDIGEI

(d)





5 Conclusions




2 4 6 8 1075

80

85

90

95

100Re

cogn

ition

accu

racy

()

Rank


(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER


(b)


5 10 15 20 25 300

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

)

Dimension

2DPCAPCA




Acknowledgments


References


































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










2 4 6 8 1075

80

85

90

95

100Re

cogn

ition

accu

racy

()

Rank


(a)

0 02 04 06 08 10

02

04

06

08

1

FAR

FRR

EER


(b)


5 10 15 20 25 300

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

)

Dimension

2DPCAPCA




Acknowledgments


References


































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation



































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









Research Article Average Gait Differential Image Based Human...

Documents

Transcript of Research Article Average Gait Differential Image Based Human...