gait fustion gait recognition based on fusion of about ...€¦ · fusion rules system overview...
TRANSCRIPT
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Gait Recognition Based on Fusion ofMulti-view Gait Sequences
Presented by Shiqi Yu
Yuan Wang1, Shiqi Yu1, Yunhong Wang2, and Tieniu Tan1
1. Institute of Automation, Chinese Academy of Sciences
2. School of Computer Science and Engineering, Beihang University
Jan. 5, 2006. Hong Kong
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Outline
1 About Gait Recognition
2 CASIA Gait Database
3 Algorithms and Fusion Rules
System Overview
Gait Recognition Algorithm
Fusion Rules
4 Experimental Results and Analysis
5 Conclusions and Future Work
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
About Gait Recognition
Gait: A particular way or manner of moving on foot.
Earlier research shows that gait can be used as abiometric for personal identification.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
About Gait Recognition
Gait: A particular way or manner of moving on foot.
Earlier research shows that gait can be used as abiometric for personal identification.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
About Gait Recognition: Advantages
Gait has several unique advantages:
Perceivable at a distance.It is the greatest difference compared with other biometrics. Itmakes gait usable in visual surveillance.Non-invasiveNon-contact
Figure: VStar visual surveillance platform
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
About Gait Recognition: Disadvantages
Disadvantages:
Gait is not stable enough, and can be influenced byweight, health condition, emotion, clothes, shoes etc.
[Chris Kirtley, Psychological Influences on Gait]
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Current Problems
Currently, most gait recognition algorithms depend on aspecific view, and can be easily affected by view, clothing,time or other factors.
Fusion of multi-view gait sequences is a solution forimproving the robustness to view change.
But this idea has not been well addressed before.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Our work
Our work focuses on
constructing a gait recognition system based onfusion of multi-view sequences,investigating which fusion rule is suitable forcombining of multi-view sequences,investigating which two views can achieve the bestperformance.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Outline
1 About Gait Recognition
2 CASIA Gait Database
3 Algorithms and Fusion Rules
System Overview
Gait Recognition Algorithm
Fusion Rules
4 Experimental Results and Analysis
5 Conclusions and Future Work
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
CASIA Gait Database
CASIA Gait Database A was created in 2001.
20 subjects3 views240 video sequences
CASIA Gait Database B was created in 2005.
124 subjects ( 94 males and 30 females)11 views3 conditions (normal walking, walking with a coat andwalking with a bag)13640 (124 × 11 × 10) video sequences
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
CASIA Gait Database
CASIA Gait Database A was created in 2001.
20 subjects3 views240 video sequences
CASIA Gait Database B was created in 2005.
124 subjects ( 94 males and 30 females)11 views3 conditions (normal walking, walking with a coat andwalking with a bag)13640 (124 × 11 × 10) video sequences
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
CASIA Gait Database B
Network
Figure: Set-up for gait data collection
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
CASIA Gait Database B
(a) Video from 11 views
(b) Normalwalking
(c) Walking witha coat
(d) Walking witha bag
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
CASIA Gait Database B
(e) Video from 11 views
(f) Normal walk-ing
(g) Walking witha coat
(h) Walking witha bag
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
CASIA Gait Database
CASIA Gait Database (subset A and B) is in the publicdomain.
For details, please visit:
http://www.cbsr.ia.ac.cn
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Outline
1 About Gait Recognition
2 CASIA Gait Database
3 Algorithms and Fusion Rules
System Overview
Gait Recognition Algorithm
Fusion Rules
4 Experimental Results and Analysis
5 Conclusions and Future Work
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
System Overview
Two views were combined to improve the performance.
Figure: Fusion system
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Gait Recognition Algorithm
The gait algorithm which we used is described in paper[Yu, ICIG’04], which is named as Key Fourier Descriptors(KFDs).
Each point on the contour can be represented as:
si = xi + j · yi
The contours in a sequence can be represented as a longvector:
[s1, s2, · · · , sN , sN+1, sN+2, · · · , s2N , · · · ]
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Gait Recognition Algorithm
The gait algorithm which we used is described in paper[Yu, ICIG’04], which is named as Key Fourier Descriptors(KFDs).
Each point on the contour can be represented as:
si = xi + j · yi
The contours in a sequence can be represented as a longvector:
[s1, s2, · · · , sN , sN+1, sN+2, · · · , s2N , · · · ]
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Gait Recognition Algorithm
We can obtain Fourier descriptors(coefficients) by discreteFourier transform(DFT)
fn =M−1∑m=0
zne−j(2π/M)nm
for n = 0, 1, 2, · · · , N − 1
Key Fourier Descriptor (KFD) feature is defined as:
F =1
|f (T )|[|f (2T )|, |f (3T )|, · · · , |f ((N − 1)T )|]
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Gait Recognition Algorithm
We can obtain Fourier descriptors(coefficients) by discreteFourier transform(DFT)
fn =M−1∑m=0
zne−j(2π/M)nm
for n = 0, 1, 2, · · · , N − 1
Key Fourier Descriptor (KFD) feature is defined as:
F =1
|f (T )|[|f (2T )|, |f (3T )|, · · · , |f ((N − 1)T )|]
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Fusion Rules
The output of N systems is X = [x1, x2, · · · , xN ].
Four fusion rules are used:
1. Sum rule
x =N∑
i=1
xi
2. Product rule
x =N∏
i=1
xi
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Fusion Rules
3. Weighted sum rule
x =N∑
i=1
wi · xi
where wi =ERR−1
iPNj=1 ERR−1
j.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Fusion Rules
4. Dempster-Shafer (D-S) ruleIn this frame of the evidence theory, the bestrepresentation of support is a belief function ratherthan a Bayesian mass distribution. D-S theoryconcerns the combination of degrees of belief and issuitable for fusion.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Outline
1 About Gait Recognition
2 CASIA Gait Database
3 Algorithms and Fusion Rules
System Overview
Gait Recognition Algorithm
Fusion Rules
4 Experimental Results and Analysis
5 Conclusions and Future Work
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Experiments
Normal walking sequences were used in the experiments.
Two views were combined to improve the performance.
There are 11 views, and totally C211 = 55 different fusion
pairs.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Experiments
Normal walking sequences were used in the experiments.Two views were combined to improve the performance.
There are 11 views, and totally C211 = 55 different fusion
pairs.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Experiments
Normal walking sequences were used in the experiments.Two views were combined to improve the performance.
There are 11 views, and totally C211 = 55 different fusion
pairs.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Experimental Results
Rule non-fusion Sum Product W-Sum D-SEER 11.49% 9.08% 8.56% 8.85% 3.81%
Table: The average EERs (Equal Error Rates) of the fusionsystems
D-S rule has the best performance.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Experimental Results
In some cases fusion system can not gain improvementcompared with the best single system.
Rule Sum Product W-Sum D-SNumber 7 6 2 0
Table: The number of experiments failed to gain improvement
The trained rules (W-Sum and D-S) are better than fixedrules (Sum and Product) in multi-view gait recognition.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Experimental Results: D-S
Improvement of the D-S fusion system
0
50
100
150
200
0
50
100
150
200−10
−8
−6
−4
−2
0
Angle 1
The improvement of fusion system
Angle 2
EE
Rfu
sion
−E
ER
min
(%)
z: the difference between the EER of fusion system andthe lower EER of single systems.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Experimental Results: D-S
Improvement of the D-S fusion system
0
50
100
150
200
0
50
100
150
200−10
−8
−6
−4
−2
0
Angle 1
The improvement of fusion system
Angle 2
EE
Rfu
sion
−E
ER
min
(%)
If the two views are respectively from range [0◦, 90◦] and[90◦, 180◦], greater improvement can be achieved.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Outline
1 About Gait Recognition
2 CASIA Gait Database
3 Algorithms and Fusion Rules
System Overview
Gait Recognition Algorithm
Fusion Rules
4 Experimental Results and Analysis
5 Conclusions and Future Work
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Conclusions
A gait recognition scheme based on fusion of multi-viewgait sequences is proposed.
Fusion can improve the performance of gaitrecognition.
Dempster-Shafer fusion rule can achieve greaterimprovement than other fusion rules.If the two views are respectively from range [0◦, 90◦]and [90◦, 180◦], greater improvement can beobtained.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Conclusions
A gait recognition scheme based on fusion of multi-viewgait sequences is proposed.
Fusion can improve the performance of gaitrecognition.Dempster-Shafer fusion rule can achieve greaterimprovement than other fusion rules.
If the two views are respectively from range [0◦, 90◦]and [90◦, 180◦], greater improvement can beobtained.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Conclusions
A gait recognition scheme based on fusion of multi-viewgait sequences is proposed.
Fusion can improve the performance of gaitrecognition.Dempster-Shafer fusion rule can achieve greaterimprovement than other fusion rules.If the two views are respectively from range [0◦, 90◦]and [90◦, 180◦], greater improvement can beobtained.
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Future Work
Our future work:
Feature level fusion of gait recognitionView invariant gait feature representation
Gait Fustion
About GaitRecognition
Database
Algorithms andFusion RulesSystem Overview
Gait RecognitionAlgorithm
Fusion Rules
Experiments
Conclusions
Shiqi YU
Center for Biometrics and Security Research (CBSR)http://www.cbsr.ia.ac.cn
National Laboratory of Pattern Recognition (NLPR),Institute of Automation, Chinese Academy of Sciences(CASIA)http://www.nlpr.ia.ac.cn
sqyu @ nlpr.ia.ac.cn