HUMAN GAIT RECOGNITION USING DIFFERENCE BETWEEN

FRAMES

Mazaher Karami and Alireza Ahmadyfard

Department of Electronics, School of Robotics and Electronic Engineering, Shahrood

University of Technology, Shahrood, Iran

Keywords: Biometric, Frieze pattern, Gait recognition.

Abstract: In this paper, we address the problem of human identification using gait. Considering the recent work of Lee

et al. (Lee et al., 2007) proposed for gait recognition. First we will introduce the algorithm proposed by Lee

et al.. This method has two main steps: (1) extract key frames to define the gait cycle pattern, and (2)

compute Shape Variation-based frieze patterns. These patterns are then used to classify and perform the gait

identification. We modify the utilized features in this approach. We try to omit redundant features based on

the effect of each feature on recognition rate and in next step, we improve performance of this approach by

making some changes in way of feature extraction. Finally, we use the statistical characteristics of employed

features instead of direct applying of remaining features. We test the proposed method on CASIA database.

The experimental results are used to compare the proposed method with Lee et al. method.

1 INTRODUCTION

Nowadays, there is an ever growing need to

determine or verify the identity of a person.

Biometrics are one of most important tools for this

purpose. Biometrics is a branch for identifying or

verifying the identity of a person based on

physiological or behavioral characteristics.

Physiological characteristics include fingerprints and

facial image. The behavioral characteristics are

actions carried out by a person in a characteristic

way and include signature, voice or gait, though

these are naturally dependent on physical

characteristics.

One of behavioral biometrics is gait. Gait as a

biometric has been attracted by many researchers in

recent years. It is non-invasively and can be

performed from distance. All human have the same

basic walking pattern, but their gaits are influenced

by many factors like their musculo-skeletal

structure, limb lengths, body mass and shape and

several other factors (Murray et al, 1964)(

Johansson,

1973

). These make gait unique for each person.

Generally, gait recognition approaches can be

categorized in two main groups: model-based

approaches and model-free or appearance-based

approaches.

Model-based methods simulate human body

using a model. In model-based approaches, a general

model is considered for human body and this model

is fitted to body of each person. Then using this

model, desired features are extracted. Model-based

approaches are fairly robust to some covariates like

view angle and occlusion, but they need large

amount of computations. Joint trajectories (Wang et

al., 2004), stride parameters (BenAbdelkader, 2002)

and so on, can be categorized in this group.

Second category of methods, are model-free or

appearance-based methods. In this category,

different algorithms are used to capture human

motion features, for example averaged silhouette

(Liu and Sarkar, 2004), HMMs (Chen et al.,

2006)(Sundaresun et al., 2003)(Kale et al.,

2002)(Suk and Sink, 2006), PCA (Murase and

Sakai, 1996), symmetry analysis (Hayfron-Acquah

et al., 2003), etc.

Lee et al. (Lee, 2007) used difference between a

key frame and the frames in a walking cycle for

feature extraction. In this paper, first we will review

their approach. Through the experiments we noticed

that some employed features are redundant, so we

omit redundant features in our proposed method.

Then we will modify algorithm to improve its

performance.

The rest of this paper is organized as follow. In

327

Karami M. and Ahmadyfard A. (2009).

HUMAN GAIT RECOGNITION USING DIFFERENCE BETWEEN FRAMES.

In Proceedings of the Fourth Inter national Conference on Computer Vision Theory and Applications, pages 327-332

DOI: 10.5220/0001793003270332

 SciTePress

section 2 we will review the algorithm proposed

by Lee et al. (Lee, 2007). In section 3 we investigate

the employed features in their algorithm and effect

of these features on recognition rate. Section 4

includes the proposed modification in order to

improve performance of Lee algorithm. Section 5

shows the experimental results and finally in Section

6 we draw the paper to conclusion.

2 PREVIOUS WORKS

Lee et al. (Lee, 2007) introduced a method for

recognition of human based on gait. The algorithm

mainly consists of two parts. First we need to extract

key frames for each gait cycle. We define one gait

cycle as the period starting from a double support

stance frames with left foot forward to the next. To

do this, we seek reliable detection of frames

occurring at the same relative offset within each gait

cycle (for example, double support stance frames

with left foot forward). Secondly, difference frames

based on subtracting these key frames from

silhouettes at other times are calculated and the

Shape Variation-Based frieze pattern (SVB frieze

pattern) is computed based on these difference

frames.

2.1 Key Frames

For computing SVB frieze pattern, we need to

determine a key frame as a reference frame for each

walking cycle. The key frame is defined as the

starting frame of one walking cycle, which is one of

the two double-support positions, (two feet on the

ground) where left foot is front. Each walking cycle

starts from the key frame and ends before the next

key frame. First, all silhouette images are aligned by

calculation centroid of silhouette.

Figure 1: Normalized distance variation vector between

feet.

To find the start point of each cycle, projection

of the lower part of body for each silhouette

perpendicular to the horizontal axis is computed and

its width is obtained. Figure 2 is the plot of this

width vector over time.

After the key frame is obtained, a series of

difference frames are computed between key frame

and successive frames in gait cycle.

),,(),,(),,( tdyydxxItyxItyxD

keykey

++−= (1)

Here, (dx,dy) is the offset for minimum frame

difference and I(x,y,t) is the frame at time t inside a

given cycle. Figure 2 shows the process of

computing difference frames based on a key frame.

2.2 Shape Variation-Based Frieze

Pattern Extraction

SVB frieze patterns can be obtained by projecting

pixel values of difference frames along horizontal or

vertical axes.

Figure 2: Key frame and difference of other frames from it

(Lee et al, 2007).

By Projection of each difference frame, it is

converted into 1D vector. The SVB frieze pattern is

obtained by putting these vectors next together over

time.

(,) (,,)

FP y t D x y t=

∑

(2)

∑

tyxDtxFP ),,(),(

(3)

),,( tyxD is difference frame at time t.

Summation in formula 2 is over x's (rows) of

),,( tyxD . In formula 3 Summation is over y's

(columns) of

),,( tyxD .

Figure 3 shows a horizontal SVB frieze pattern

and Figure 4 is a vertical SVB frieze pattern. Each

column of a SVB frieze pattern at time t represents

the difference frame at time t.

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

328

Figure 3: Typical horizontal SVB frieze pattern.

Figure 4: Typical vertical SVB frieze pattern.

2.3 Symmetry Map of Frieze Pattern

One gait cycle consists of two half cycles. These two

half cycles have almost the same pattern. But these

two half cycle are not exactly similar. In fact there is

some differences between them. These differences

can be used as a feature for recognition. In Lee

algorithm, these differences between two half cycles

has been obtained from SVB frieze patterns by

computing the difference between two half motion

cycles of SVB frieze patterns as symmetry map

(SM).

2.4 Classification

Four cues from each gait sequence: horizontal &

vertical SVB frieze pattern and horizontal and

vertical symmetry map. To obtain distance between

i’th gallery to j’th probe following distance are

computed:

hFP

FPFP

−=Φ

(4)

vFP

FPFP

−=Φ

(5)

hSM

SMSM

−=Φ

(6)

vSMv

SMSM −=Φ

(7)

A single cost function is computed by summing up

all four distance values.

3 EFFECT OF EACH FEATURE

ON RECOGNITION RATE

In order to study the effect of each feature on

recognition rate, we implemented the algorithm for

individual features once at a time. We also

evaluated the algorithm when SMh and SMv or FPv

and FPv are employed as a pair. Finally recognition

rate is computed when all four features are used all

together.

We tested the algorithm on CASIA database.

Obtained results in table 1 show that SMv and SMh

have least success in making distinction between

individuals. Best results achieved when FPh and

FPv are used together. When all four features are

utilized for identification, recognition rate is less

than while only FPh with FPv are used as a pair.

This indicates that SMv and SMh not only do not

increase recognition rate but also degrade it.

Therefore, we suggested using only FPh and FPv for

our algorithm.

Table 1: Ability of each feature in recognizing individuals.

4 PROPOSED CHANGES TO

IMPROVE PERFORMANCE

In this section we introduce the modification

suggested to improve Lee et al. method (Lee, 2007).

First we use difference between successive frames

rather than key frame based method, used in original

paper. The experimental results show that this

modification improves the recognition performance.

In next step, we try out statistical characteristics of

frieze patterns instead of direct use of them for

calculation of cost function.

4.1 Difference Frames

In Lee et al. algorithm difference between a key

frame with frames in a walking cycle is used to

nm-05 nm-06

Rank 1 Rank 5 Rank 1 Rank

FPh 62.2% 76.3% 61% 80.5%

FPv 68% 85.7% 72.9% 82.2%

SMh 46.2% 59.6% 46.6% 53.4%

SMv 44.5% 58% 42.3% 60.2%

FPh + FPv 68.9% 85.7% 71.2% 83.9%

SMh + SMv 47% 61.3% 50% 57.6%

All four

features

64.7% 84% 67.8% 81.3%

HUMAN GAIT RECOGNITION USING DIFFERENCE BETWEEN FRAMES

329

compute SVB frieze pattern. We can calculate

difference between successive frames instead of

difference between a key frame and other frames in

a walking cycle. Figure 5 shows successive

difference frames.

Figure 5: Successive difference between frames.

Difference between frames can be a

representative for motion in a way and using this, we

can capture dynamic features of walking. For this

purpose, we obtain a starting frame for each gait

cycle as previous. Then, differences between

consecutive frames are achieved. By means of these

new difference frames, FPh and FPv are computed.

Table 2 indicates achieved results. By implying

this change to algorithm, performance fairly

improves.

4.2 Statistical Characteristics of Frieze

Pattern

Direct use of FPh and FPv in order to computation

of distance between features of two individuals has

some disadvantages. It increases sensitivity to noise.

Moreover, the length of features must be the same to

make comparison possible. So we need to align

them, and this also adds extra noise to system. We

can extract some characteristics of these features,

and use these characteristics for calculation of

distance function, instead of direct use of features.

Calculation of statistical moments is one way to

capture embedded information in given data. Mean

and variance are most popular statistical moments.

For this purpose, we calculate mean value of data in

each row of FPh and FPv. By doing this, we convert

each nm× matrix to a vector with size of

m .

We apply this algorithm to FPh and FPv. Now,

these mean vectors are used for obtaining distance

functions rather than FPh and FPv themselves. We

name these new features as MFPh and MFPv. By

means of this change, we will have a great increase

in recognition rate.

5 EXPERIMENTAL RESULTS

In this section, we demonstrate the result of

experiment in which our algorithm is compared with

the Lee et al. algorithm (Lee et al, 2007). In Lee et

al. paper, MoBo database is used for evaluating the

algorithm. We did not access to MoBo database

instead we used CASIA database in this experiment.

Nearest neighbour method is used for classification.

5.1 CASIA Database

CASIA gait database is collected by Institute of

Automation, Chinese Academy of Sciences. This

database is available on "http://www.sinobiometrics.

com". CASIA database contains 124 subjects and for

each subject there are 10 different capturing

conditions ("nm-01" to "nm-06","bg-01","bg-

02","cl-01" and "cl-02"). In subject nm-01 to nm-06,

person is walking freely in different times but

without changes in its appearance. In bg-01 and 2,

subject is carrying a bag and in cl-01 and 2, clothing

is changed. For each of above conditions subject is

viewed from 11 different angles. We used only nm-

01 to nm-06, while the viewing angle for walking

subject is perpendicular to optical axis of camera.

We used subsets nm-01 to 4 for training and nm-05

and nm-06 subsets for test.

5.2 Results

We have tabled the result of implementing original

algorithm proposed by Lee et al. and our proposed

changes in section 4 on CASIA database in tables 2

and 3, respectively.

Table 2 is result of algorithm after use of

successive difference frames instead of key frame

based difference frames. Fair improvement is

observed by applying this change on main work. In

this stage as a result of previous discussion about

effect of each feature on recognition rate we ignored

SMh and SMv and only FPh and FPv have been

computed. First row shows recognition rate, when

difference frames have been computed based on

difference between a key frame and the frames in

other times. Second row shows result of algorithm

when successive difference frames have been used

instead of key frame based differences.

Table 3 demonstrates achieved result using rows

mean vectors (MFPh and MFPv) in comparison with

direct application of FPh and FPv. Here, for

computing the FPh and FPv, successive difference

frames have been used. In order to use mean vectors,

we first normalized each vector by subtracting

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

330

Table 2: Result of using successive difference between

frames and computing FPh and FPv.

nm-05 nm-06

Rank 1 Rank 5 Rank

Rank

Key frame

differences

68.9% 85.7% 71.2% 83.9%

Successive

differences

74.8% 84% 73.7% 84.7%

mean of each vector from it and then dividing it by

its standard deviation. We obtain considerable

increase in recognition rate when we use mean

vectors.

Table 3: Result of using row mean of frieze patterns

instead of frieze patterns themselves.

6 CONCLUSIONS

In this paper, we introduced one way to recognize

people based on their gait, proposed S. Lee et al.

from the Penn state university. We tried to omit

redundant used features in this algorithm. Then we

applied differences between consecutive images to

extract features instead of computation of difference

between a key frame and other frames. Using these

frames, vertical and horizontal frieze patterns are

computed. In calculation of distance function, mean

value of each row of frieze patterns in form of a

vertical mean vector and a horizontal mean vector

are used. We showed that applying mean vectors is

more successful than direct use of frieze patterns.

We implemented our algorithm and previous

work, on CASIA database. We indicated that our

algorithm has better performance in comparison.

ACKNOWLEDGEMENTS

The authors thank Chinese Academy of Sciences.

Portions of the research in this paper use the CASIA

Gait Database collected by Institute of Automation,

Chinese Academy of Sciences.

REFERENCES

Z. Liu, S. Sarkar, 2004, "Simplest representation yet for

gait recognition: averaged silhouette", Proceedings of

the 17th International Conference on Pattern

Recognition, Vol. 2,704-711.

C. Chen, J. Liang, H. Zhao, H. Hu, 2006, "Gait

recognition using hidden markov model", ICNC, Part

I, LNCS 4221, pp.399-407, Springer.

A. Sundaresun, A. R. Chowdhury, R. Chellappa, 2003, "A

hidden markov model based framewok for recognition

of humans from gait sequences", Proceeding of IEEE

International Conference on Image Processing.

A. Kale, A. Rajagopalan, N. Cuntoor, and V. Kruger,

2002, "Gait-based recognition of humans using

continuous HMMs", Proceeding of the IEEE

Conference on Automatic Face and Gesture

Recognition.

H.I. Suk, B.K. Sink, 2006, "HMM-based gait recognition

with human profiles", SSPR and SPR, LNCS 4109,

pp.596-603.

H. Murase, R. Sakai, 1996, "Moving object recognition in

eigenspace representation: gait analysis and lip

reading", Pattern Recognition Lett.17, 155–162.

J. B. Hayfron-Acquah, M. S. Nixon, J. N. Carter, 2003,

"Automatic gait recognition by symmetry analysis",

Pattern Recognition Letters, ELSEVIER Science B.V.

S. Lee, Y. Liu, and R. Collins, 2007 , "Shape Variation-

based Frieze Pattern for Robust Gait Recognition,"

Proceedings of CVPR 2007, 1-8.

M. S. Nixon, J. Carter, 2006, "Automatic Gait Recognition

by Gait", Proceedings of the IEEE, Vol. 94, No. 11.

CASIA Gait Database,

http://www.sinobiometrics.com

M.P. Murray, A.B. Drought, R.C. kory, 1964, "Walking

Pattern of Normal Men", J. Bone Joint Surg. 46 A(2)

33-360.

G. Johansson, 1973 “Visual Perception of Biological

Motion and a Model for its Analysis”, Perception and

Psychophysics, Vol. 14, No. 2, pp. 201–211.

P.S. Huang, C.J. Harris, M.S. Nixon, 1999, "Human Gait

Recognition in Canonical Space Using Temporal

Templates", IEE Proceeding Computer Vision, Image

& Signal Processing.

L. wang, T. Tan, 2003, "Silhouette Analysis-Base Gait

Recognition for Human Identification", IEEE Trans.

PAMI, 25(12) 1505-1518.

L. Wang,T. Tan, H. Ning, W. Hu, 2004, "Fusion of Static

and Dynamic Body Biometrics for Gait Recognition",

IEEE Trans. Circuit Sys. Video Techno., vol. 14, no.

2, pp. 149-158.

C. BenAbdelkader, 2002, "Gait as a Biometric for Person

Identification in Video Sequences", PhD Dissertation,

University of Maryland, College Park.

nm-05 nm-06

Rank 1 Rank 5 Rank 1

Rank

FPh + FPv 74.8% 84% 73.7% 84.7%

MFPh+MFPv 90.7% 95% 87.3% 95.8%

HUMAN GAIT RECOGNITION USING DIFFERENCE BETWEEN FRAMES

331