gait recognition by deformable registration · 6/18/2018 · gait recognition by deformable...
TRANSCRIPT
Gait Recognitionby Deformable Registration
Yasushi Makihara1, Daisuke Adachi1, Chi Xu2,1, Yasushi Yagi1
1: The Institute of Scientific and Industrial Research, Osaka Univ.2: School of Computer Science and Technology, Nanjing University of Science and Technology
1
IEEE Computer Society Workshop on Biometrics 2018, Jun. 18th, 2018
Gait recognition: Overview
2Distance from sensor
DNA FaceFinger print IrisVein Gait
Near Far
Recognition at a distanceCriminal investigation
Gait recognition: Use casesAdmitted as evidence in courts for the first time UK in 2008 Japan in 2016
3
Gait verification systemfor criminal investigation [Iwama+ 2013]
[1] http://news.bbc.co.uk/2/hi/programmes/click online/7702065.stm, ”How biometrics could change security,” BBC News, 31 Oct. 2008.
Challenge of gait recognition
4Challenge: Intra-subject posture changes
Distracted walking is often seen in the real world.
Probe(inclined forward)
Gallery(normal)
Largedissimilarity
Falselyrejected
Related work
5
Combination of appearance-basedgait feature and metric learning
RankSVM[Martin-Felez+ PR2014]
Deformable registration model for gait recognition is required.
Robust gait recognition
Gait energy image (GEI)[Han+ TPAMI2006]
+
Not direct way to handle geometric deformation such as posture change
Registration model for face
w/ active shape model (ASM) [Thai+ IJBB2011]
Expression-invariant face recognition
w/ robust constrained local models [Boddeti+ 2017]
Outstanding landmarks such as eyes, nose, mouth, are unavailable for gait
Objective Gait recognition by deformable registration
6
Probe(Inclined forward)
Gallery(Normal)
Dissimilarity(Large)
Falselyrejected
Probe(Deformed)
Deformable registration model
Trulyaccepted
Dissimilarity(Small)
Free-form deformation (FFD)
Deformation vector on the control points (CPs)
Overall deformation field:Bilinear interpolation from adjacent CPs
7
Probe image Ip Deformed probe image F(Ip; u)
ui,j uij
Deformable registration
pi,jpi,j pijuijpi,jui,j
ui,j pi,j pijuij
Computing FFD between two images
8
Intra-/inter-subject FFDs
9
Intra-subject
Inter-subject
Forward inclination
Body shape
Translation
Head-torso ratio Arm swing Stride
Different trend Learn intra-subject deformation
# Deformation is represented by morphing.
Eigen FFD
10
Matching by eigen FFD
11
Metric learningConvolutional neural network:Matching local features at the bottom layer [Wu+ TPAMI 2016]
12
128
W1
16
122
W1’
16
122
128
Def
orm
ed p
robe
Gal
lery
16
122
W3
7
256
16
61
W2
Black digits: SizeBlue digits: #Channels
Green digits: Stride: Pooling
: Convolution filter
: Normalization
64
55 28
64
7 16
7 16
7 64 2 22 2256
22
W4
: Full connection w/ dropout
01
Experimental setup
13
Matching example
14
(a) Probe(inclined forward)
Fals
e m
atch
True
mat
ch
(b) Gallery
Dissimilarity score
1.42
1.31
0.60
Difference (a) – (b)
1.11
(a’) Deformed probe
Difference (a’) – (b)
(a) Probe(inclined forward)
Evaluation w/o metric learning
15
Receiver operating characteristics(ROC) curves
Better
Cumulative matching characteristics(CMC) curves
Better
False acceptance rate (FAR)
Fals
e re
ject
ion
rate
(FR
R)
Evaluation w/ metric learning
16
ROC curves CMC curves
Better
Better
Fals
e re
ject
ion
rate
(FR
R)
False acceptance rate (FAR)
Summary Gait recognition by deformable registration robust
against intra-subject posture change
Future work Handle more posture changes (e.g., climbing up stairs) Jointly optimize deformable registration and metric
learning 17
World’s largest gait database
18
Data set #Subjects Covariates
OUMVLP[Takemura+2017]
10,307
14 views
OULP-Bag[Uddin + 2018] 62,528
Carried objects in the wild
OULP-Age[Xu+ 2018] 63,846
Wide age range
World’s largest gait database Available at
http://www.am.sanken.osaka-u.ac.jp/BiometricDB/index.html
19