http://isrc.ulster.ac.uk p rw gei: poisson random walk based gait recognition intelligent systems...
TRANSCRIPT
http://isrc.ulster.ac.uk
PRWGEI: Poisson Random Walk based Gait Recognition
Intelligent Systems Research CentreSchool of Computing and Intelligent Systems,
University of Ulster, Northern Ireland, UK.
1
Pratheepan Yogarajah, Joan V. Condell, Girijesh Prasad
http://isrc.ulster.ac.uk
The Problem statement Existing methods Our solution Experimental Results Summary
2
Overview
http://isrc.ulster.ac.uk
The Problem Performance of Gait Recognition under different
Covariate Conditions.● What?
– Gait recognition is recognizing people by the way they walk.
● Why?– Gait recognition is non intrusive– Operates at a distance without subject cooperation.
3
http://isrc.ulster.ac.uk4
Covariate Conditions?• Conditions that effect gait.• Can be divided into two categories
1) Effecting features extracted from gait- Carrying Condition, Clothing Condition, View etc
2) Effecting gait itself- Shoes, Time, Injury, Speed etc
Normal Carrying Condition Clothing Condition
http://isrc.ulster.ac.uk5
Gait feature representation
1. Appearance-Based Methods: Appearance-based approaches are widely used in gait
representation. They directly represent human motion using image information, such as a silhouette, an edge, and an optical flow.
2. Model-Based Methods: Model-based approaches represent a gait with body
segments, joint positions, or pose parameters.
http://isrc.ulster.ac.uk6
Existing appearance-based works
GEIPAMI 06
EGEISP 08
GEnIPRL 10
MG
PRL 10AEI
SP 10
• Even though these gait feature representations show good recognition rate, their average recognition rates are not that promising (i.e. less than 75%). • This indicates that a more robust appearance-based gait feature representation is needed.
http://isrc.ulster.ac.uk7
Our Solution
• Parts structure based GEI (PRWGEI).• Apply Poisson Random Walk approach to
binary silhouette.• Extract various properties of shape analysis
Binaryimage
Transformedimage
http://isrc.ulster.ac.uk8
Poisson Equation
http://isrc.ulster.ac.uk9
Random walk
Mean time to
hit the boundary
http://isrc.ulster.ac.uk10
Poisson based shape representation
http://isrc.ulster.ac.uk11
PRWGEI
The columns from left to right represent PRWGEI features for normal, carrying objects and different clothing covariate factors.
http://isrc.ulster.ac.uk12
Recognition
• The PCA+LDA is applied to reduce the dimension of the data and also to improve the discriminative power of the extracted features.
• Then the k-nearest neighbour (k-NN) is applied to classify the data and make a decision, i.e. decide the identity of a person.
http://isrc.ulster.ac.uk13
Subspace of dimensionality - (PCA)
140 provides better recognition rate
http://isrc.ulster.ac.uk14
Results (CASIA-B)
GEI GEnI MG AEI PRWGEI
CasiaSetA2 99.4% 98.3% 100% 88.7% 98.4%
CasiaSetB 60.2% 80.1% 78.3% 75.0% 93.1%
CasiaSetC 30.0% 33.5% 44.0% 57.3% 44.4%
Overall 63.2% 70.6% 74.1% 73.7% 78.6%
• 124 individuals with normal(6), carrying(2) and clothing(2) conditions.• training : 124 individuals with normal(4) – CasiaSetA1• testing : 124 individuals with normal(2) – CasiaSetA2, carrying(2) – CasiaSetB and clothing(2) - CasiaSetC
http://isrc.ulster.ac.uk15
Summary• Gait Energy Image representation based on
Poisson random walk approach
• Our proposed novel gait features provide better results for person identification with CASIA-B dataset.
• As a future direction, we would like to test our method with dataset such as USF dataset and SOTON dataset.
http://isrc.ulster.ac.uk
Thank you
16