Upload
vanngoc
View
218
Download
1
Embed Size (px)
Citation preview
대한전자공학회 2002 하계종합학술대회Dynamic Footprint-based Person Recognition Method-1/59
Dynamic Footprint-basedPerson Recognition Method and
Application to Intelligent Residential Space
August 25, 2004
Jin-Woo Jung
Human-friendly Welfare Robot System Research CenterKAIST, KOREA
1st UNSW-KAIST joint seminar
Dynamic Footprint-based User Identification System-2/59
ContentsI. Introduction
II. Literature Review: Gait vs. Footprint1. Gait vs. Footprint2. Problem Statement
III. Dynamic Footprint1. Definition of dynamic footprint2. Measurement of dynamic footprint3. Footprint database using MAT sensor (MAT-DB)
IV. Dynamic Footprint-based User Identification System1. User Identification Method via Overlapped Footprint Shape2. User Identification Method via COP Trajectory3. Information fusion
V. Concluding Remarks
대한전자공학회 2002 하계종합학술대회Dynamic Footprint-based Person Recognition Method-3/59
I. Introduction
Dynamic Footprint-based User Identification System-4/59
Customer’sBehaviors
Observation
Modeling
Left-handedor
Right-handed
Personalized Service : A veteran waitress in a restaurant
Introduction
Dynamic Footprint-based User Identification System-5/59
Personalized Service : A short-cut from the movie “Minority Report”
• A Person comes in the GAP store.• The system tells him with his biometrics and his
previous purchase data as follows:– Hello, Mr. Yakamodo.– Welcome back to the GAP.– How was the tank-tops of GAP for you?
Introduction
Dynamic Footprint-based User Identification System-6/59
Realize a personalized service for multi-user living environmentby human-friendly user identification method !
User1User2
…UserN
UserIdentification
Multi agentLearningSystem
IntentionReading
VariousEnvironmental
SensorsVariousHome
Appliances
Motivation: Personalized serviceIntroduction
Dynamic Footprint-based User Identification System-7/59
Historical Trends of User Identification
• “something that you possess”– Lost, stolen, forgotten, misplaced, or abused
• “something that you know” & “something that you possess”– Difficult to remember, easily recallable password
• “something that you are” (biometrics)– Cannot be misplaced or forgotten
[Jain 1999] A. Jain, R. Bolle, and S. Pankanti, Biometrics: Personal Identification in Networked Society, KAP press, 1999
Introduction
To enhance security
To enhance convenience(human-friendliness)
Dynamic Footprint-based User Identification System-8/59
Remark: What is biometrics?• A science involving the statistical analysis of
biological characteristics (wide-sense)
• The automated method of physiological or behavioral characteristics to determine or verify identity (narrow-sense)
– Universality– Uniqueness– Permanence– Collectability– Performance– Acceptability (Human-friendliness)– Circumvention
[Zhang 2000] D. D. Zhang , Automated Biometrics – Technologies and Systems, KAP press, 2000
Introduction
Dynamic Footprint-based User Identification System-9/59
Comparison of biometrics
[Jain 1999] A. Jain, R. Bolle, and S. Pankanti, Biometrics: Personal Identification in Networked Society, KAP press, 1999
0.5101000.5BehavioralGait
0100.5000.5BehavioralVoice
0101000BehavioralSignature
0.50.500.5000BehavioralKeystrokes
0.510.50.510.50.5PhysiologicalEar
0010111PhysiologicalDNA
00.500111PhysiologicalOdor
110.51011PhysiologicalThermograms
10100.511PhysiologicalRetina
1010.5111PhysiologicalIris
10.50.50.50.50.50.5PhysiologicalHand Vein
0.50.50.510.50.50.5PhysiologicalHand
Geometry
10.510.5110.5PhysiologicalFingerprint
01010.501PhysiologicalFace
Circumvention
Acceptability
Performance
Collectability
Permanence
Uniqueness
Universality
CharacteristicBiometrics
1: High, 0.5: Medium, 0: Low
Introduction
Dynamic Footprint-based User Identification System-10/59
0
0.2
0.4
0.6
0.8
1
1.2
Unive
rsa l
ityUni
quen
ess
Perm
anen
ceCol
lect
abilit
yPe
rform
ance
Acce
ptab
il ity
Circum
vent
ion
Fingerprint
Iris
0
0.2
0.4
0.6
0.8
1
1.2
Unive
rsa l
ityUni
quen
ess
Perm
anen
ceCol
lect
abilit
yPe
rform
ance
Acce
ptab
il ity
Circum
vent
ion
S ignature
Voice
Gait
[Jain 1999] A. Jain, R. Bolle, and S. Pankanti, Biometrics: Personal Identification in Networked Society, KAP press, 1999
Physiological vs. Behavioral
Note: behaviroal/physiological distinction is slightly artificial.An important characteristic is that behavioral biometrics has time series data.
High accuracy-oriented application
Human-friendlyapplication
Introduction
Dynamic Footprint-based User Identification System-11/59
Behavioral biometrics
Speaker Recognition
Signature System
Keystroke biometrics
Gait biometrics
[Zhang 2000] D. D. Zhang , Automated Biometrics – Technologies and Systems, KAP press, 2000
Microphone
Electric Pad like PDA
Keyboard
Camera (Indoor/Outdoor)
Gait information is one of the most appropriate methodfor the general living environment including futuristic home environment !
Introduction
Requirement onthe environment
Behavioral biometrics[Zhang 2000]
Intentionality
Collectability Acceptability
Predeterminedsentence
Predeterminedwriting
Natural stroking
Natural walking
대한전자공학회 2002 하계종합학술대회Dynamic Footprint-based Person Recognition Method-12/59
II-1. Gait vs. FootprintII-2. Problem Statement
II. Literature Review andProblem Statement
Dynamic Footprint-based User Identification System-13/59
Gait as “a total walking cycle”• M. P. Murray et al. (1967) suggests that gait could be considered as a
total walking cycle (periodic) and, if all gait movements were considered, gait is unique.
• Recent research trends on gait recognition
Literature Review: Gait vs. Footprint
Template-based approach [Huang 1999]Model-based approach [Jain 1999]
[Jain 1999] A. Jain, R. Bolle, and S. Pankanti, Biometrics: Personal Identification in Networked Society, KAP press, 1999[Huang 1999] P. S. Huang, C. J. Harris and M. S. Nixon, “Human gait recognition in canonical space using temporal
templates”, IEE Proc. of Vision, Image and Signal Processing, Vol. 146, No. 2, 1999
Dynamic Footprint-based User Identification System-14/59
[Yam 2001] C.-Y. Yam, M. S. Nixon, and J. N. Carter, “Extended Model-Based Automatic Gait Recognition of Walkingand Running”, 3rd Proc. Audio- and Video-based Biometric Person Authentication, pp.284-294, 2001
[Yam 2002] C.-Y. Yam, M. S. Nixon, and J. N. Carter, “Performance Analysis on New Biometric Gait Motion Model”,Proc. of 5th IEEE Southwest Symposium on Image Analysis and Interpretation, pp. 31-34,, 2002
[Yam 2002]
[Yam 2001]
Report
91.7%84.2%20
92%96%5
Recog. Rate of running person
Recog. rate of walking person
Num. of person
Issues in gait recognition: different speed
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-15/59
Issues in gait recognition: different viewing direction
[BenAbdelkader 2002] C. BenAbdelkader, R. Cutler, and L. Davis, „View-invariant Estimation of Height andStride for Gait Recognition“, Lecture notes in computer science, No. 2338, pp.155-167, 2002
65 %17Non-fronto parallel
47 %41Fronto-parallel
Recognition ratesNumber of personsDirection of view
Fronto-parallelview
Non-frontoParallel view
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-16/59
• Multi-modal biometrics
[Shakhnarovich 2001] G. Shakhnarovich, L. Lee, T. Darrell, “Integrated Face and Gait Recognition from Multiple Views”, Proc. of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Vol.1, pp.439-446 , 2001
[Shakhnarovich 2001]
Issues in gait recognition: Low performance
Face + Gait
Using multiple views
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-17/59
Age, gender, race, shoes andage at which wearing shoes begins
Different foot shape
[Razeghi 2002] M. Razeghi and M. E. Batt, “Foot type classification: a critical review of current methods”,Gait and Posture, Vol. 15, pp. 282-291, 2002
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-18/59
Human footprint as a biometrics
[Kennedy 1996] R. B. Kennedy, “Uniqueness of bare feet and its use as a possible means of identification”,Forensic Science International, Vol. 82, Issue 1, pp.81-87, 1996
• R. B. Kennedy suggests that footprint is unique in each person by comparing 38 measurements from 3000 people’s 6000 inked barefoot impression data.
• With 5mm margin, there was no same foot.
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-19/59
Footprint feature used by Kennedy’s work
[Kennedy 1996] R. B. Kennedy, “Uniqueness of bare feet and its use as a possible means of identification”,Forensic Science International, Vol. 82, Issue 1, pp.81-87, 1996
bordercenterLMT
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-20/59
Automated footprint recognition
Assumption: Same body posture at every measurement time
Result: 85 % (10 men, 10 training images per individual)
[Nakajima 2000] K. Nakajima, Y. Mizukami, K. Tanaka, T. Tamura, “Footprint-based Personal Recognition”, IEEE Tr. on Biomedical Engineering, Vol. 47, No. 11, Nov. 2000
• K. Nakajima suggests the first version of automated footprint recognition system based on template matching, COP distance and foot angle.
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-21/59
Features used by Nakajima
[Nakajima 2000] K. Nakajima, Y. Mizukami, K. Tanaka, T. Tamura, “Footprint-based Personal Recognition”, IEEE Tr. on Biomedical Engineering, Vol. 47, No. 11, Nov. 2000
Normalized footfor template matching
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-22/59
Human footprint as a biometrics
Kennedy’s work (1996) Nakajima’s work (2001)
• R. B. Kennedy suggests that footprint is unique in each person by comparing 38 measurements from 3000 people’s 6000 inked barefootimpression data.
• Result:, no same foot with 5mm margin (3000 people)
• K. Nakajima suggests the first version of automated footprint recognition system based on template matching, COP distance and foot angle.
• Result: 86.55 % (10 men, 11 training images per individual)
[Kennedy 1996] R. B. Kennedy, “Uniqueness of bare feet and its use as a possible means of identification”,Forensic Science International, Vol. 82, Issue 1, pp.81-87, 1996
[Nakajima 2001] K. Nakajima, Y. Mizukami, K. Tanaka, and T. Tamura, “A New Biometrics Using Footprint”,Transactions-Institute of Electrical Engineers of Japan D, Vol.121, No.7, pp.770-776, 2001
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-23/59
Comparison of Kennedy’s and Nakajima’s work: spatial resolution
Kennedy’s data Nakajima’s dataSensor size = 10 x 10 mm2
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-24/59
Comparison: overlapped footprint during walking vs. standing footprint
Kennedy’s work Nakajima’s work
• Footprints were taken under controller conditions ensuring that each person steps onto the paper in the same manner.
• Footprints were taken under controller conditions ensuring that each person stood straightat the center of mat and looked toward a mark on the wall. And then averaging data during 5 sec.
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-25/59
Comparison: local feature vs. global feature (template matching)
Kennedy’s work Nakajima’s work
Literature Review: Gait vs. Footprint
Dynamic Footprint-based User Identification System-26/59
Problem Statement
Fusion of footprint and gait information
Vision(lateral-view)
Pressure sensing mat
MeasurementDevice
84.2% (walking)91.7% (running)
(20 person, 6 samples)
86.55%(10 person, 11 samples)
RecognitionResult
Disadvantages
View-dependency(Non-lateral view: 65% in 17 person,
7 samples [BenAbdelKader 2002])
Lighting, Clothing,Occlusion, Privacy
Gait-based[Yam 2002]
No shoes,High price (more than $15000),
Weight-loading
Footprint-based
[Nakajima 2001]
[Nakajima 2001] K. Nakajima, Y. Mizukami, K. Tanaka, and T. Tamura, “A New Biometrics Using Footprint”,Transactions-Institute of Electrical Engineers of Japan D, Vol.121, No.7, pp.770-776, 2001
[Yam 2002] C.-Y. Yam, M. S. Nixon, and J. N. Carter, “Performance Analysis on New Biometric Gait Motion Model”,Proc. of 5th IEEE Southwest Symposium on Image Analysis and Interpretation, pp. 31-34,, 2002
[BenAbdelKader 2002] C. BenAbdelkader, R. Cutler, and L. Davis, “View-invariant Estimation of Height and Stridefor Gait Recognition”, Lecture notes in computer science, No. 2338, pp.155-167, 20021
Problem Statement
Dynamic Footprint-based User Identification System-27/59
Problem Statement“Given N person’s sequential walking footprints and a new sequential walking footprint M, find the person whose walking footprint is most similar to the walking footprint M.”
Problem Statement
Sequential walking footprint (my idea !)= 2D projected data of 3D human gait
= Dynamic information (gait)Static information (foot shape)
COP/COA trajectoryOverlapped footprint
Human walking Sequential walking footprint (dynamic footprint)
Foot shape
COPtrajectory
대한전자공학회 2002 하계종합학술대회Dynamic Footprint-based Person Recognition Method-28/59
III. Dynamic Footprint
III-1. Definition of dynamic footprintIII-2. Measurement of dynamic footprintIII-3. Footprint database using MAT sensor
Dynamic Footprint-based User Identification System-29/59
Characteristics of sequential footprints from human walking
I. Repeatability
“Gait could be considered as a periodic walking cycle.” [Jain 1999]Footprints during walking are 2D projected information of human gait.
Therefore, footprints during walking also have a repeatability.
II. Asymmetry
“The shape of the left foot is typically different fromthat of the right foot.” [Sforza 1998]
Footprint patterns of each foot during walking are also asymmetric.
Dynamic footprint
[Jain 1999] A. Jain, R. Bolle, and S. Pankanti, Biometrics: Personal Identification in Networked Society, KAP press, 199[Sforza 1998] C. Sforza, G. Michielon, N. Fragnito, and V. F. Ferrario, "Foot Asymmetry in Healthy Adults:
Elliptic Fourier Analysis of Standardized Footprints", Jr. of orthopedic research, Vol.16, No.6, pp.758-765, 1998
Dynamic Footprint-based User Identification System-30/59
Definition of dynamic footprint
Dynamic footprint
A (discrete-)time sequence of footprint images FK , K=0, 1,···, T-1
is called a dynamic footprint if the following conditions are satisfied:
1) ∃ m ∈ N s. t. FK ≈ Fm+K for ∀ K=0,…,T-1
(“A ≈ B” means A=R (T (B)) for some rotation R and Translation T)
2) F 0 is the image at the right or left heel
contact time in a single gait cycle.
For K=0, 1,···, T-1, FK : [0≤x≤Wx, 0≤y≤Wy] [0≤p≤Pmax](Wx/Wy: spatial width/height of sensor, Pmax: normalized max. pressure value)
F 0 F 0
F 0 F T-1
Dynamic Footprint-based User Identification System-31/59
Definition of dynamic footprint
Dynamic footprint
Overlapped Footprint (Shape): F (x, y) = 1 if F k (x, y) ≥ Pthreshold
0 otherwise
where Pthreshold is the threshold for noise elimination
Center-Of-Pressure Point in F k: COPk = (xK*, yK*)
where xK*= , yK*=
Center-Of-Pressure Trajectory: COPT = {COP0, COP1, COP2,…,COPT-1}
For K=0, 1,···, T-1, FK : [0≤x≤Wx, 0≤y≤Wy] [0≤p≤Pmax](Wx/Wy: spatial width/height of sensor, Pmax: normalized max. pressure value)
U1
0
−
=
T
K
∑∑ ∑
yxK
x yK
yxF
yxFx
,
),(
),(
∑∑ ∑
yxK
y xK
yxF
yxFy
,
),(
),(
F 0 F T-1
Dynamic Footprint-based User Identification System-32/59
Measurement of dynamic footprint
Dynamic footprint
humanfoot
floor
Location of sensor
MAT sensor Shoe sensor
Noise sensitivity
Human’s burden
Action segmentation
Feature extraction
MAT sensor
Low
Low
Easy
Difficult
Shoe sensor
High
High
Difficult
Easy
ResponseTime
Sensitivity
Pressuresensor
Fast
High
Temperaturesensor
Slow
High
Mat-type pressure sensor[REF] J.-W. Jung, T. Sato and Z. Bien, "Dynamic Footprint-based Person Recognition Method using Hidden Markov Model
and Neural Network", Accepted for publication in Int. Jr. of Intelligent Systems
Dynamic Footprint-based User Identification System-33/59
Assumptions on measurement
Dynamic footprint
(A-1) There is no other object on the MAT sensor except the walking person.
(A-2) There is only one walking person on the MAT sensor in each trial.
(A-3) During one’s walking on the MAT sensor, at least one element ofthe MAT sensor is fired.
(A-4) Normally, user always walks along the designated direction(for example, outdoor-to-indoor) on the MAT sensor.
To find the starting time of dynamic footprint
To find the ending time of dynamic footprint
To exclude the change of walking direction
Dynamic Footprint-based User Identification System-34/59
Assumptions on measurement (cont’d)
Dynamic footprint
(A-5) User’s step length (the distance between the COA point of the first footand that of the second foot) is greater than a predefined maximum footlength, LMAX_FOOT.
(A-6) The ‘two’ footprints are always acquired after a single gait cycle.
(A-7) The user has to walk on the MAT sensor without wearing any shoes(that is, barefoot walking is assumed).
To discriminate the first foot and second foot parts
To exclude the imperfect footprint
To exclude the effect of the sole of shoes
Dynamic Footprint-based User Identification System-35/59
Assumptions on walking person
Dynamic footprint
(A-8) The person who may sometimes drag their foot after another isexcluded from the subjects.
(A-9) Users are in normal mood when they walk.
(A-10) Users have not any additional loading weight except their bodyand clothes.
To exclude abnormally walking users (users with foot pain)
To exclude the emotional effect
To exclude the effect of weight loading
Dynamic Footprint-based User Identification System-36/59
Footprint database, MAT-DB
Footprint DB usingmat-type sensor
• Number of person = 11 (M:9, F:2)• Number of data per each person
= 40• Period of data gathering = 40 days• Contents of data = dynamic footprint• Sensor = FOOT ANALZER• Frame rate = 30 Hz
Dynamic footprint
FOOT ANALZER (TechStorm, Inc.)
-Total size: 400 x 800 mm2- Spatial resolution: 10 x 10 mm- Number of sensors: 40 x 80- Max. sampling speed = 30 Hz- FSR-based mat-type pressure sensor
대한전자공학회 2002 하계종합학술대회Dynamic Footprint-based Person Recognition Method-37/59
IV-1. User Identification Method via Overlapped Footprint ShapeIV-2. User Identification Method via Center-Of-Pressure TrajectoryIV-3. Information fusion
IV. Dynamic Footprint-based User Identification System
Dynamic Footprint-based User Identification System-38/59
Feature Extraction
WalkingFootprint
OverlappedFootprint
ShapeExtraction
User Identification
Footprint ShapeMatching(Type I)
UserID
RegisteredFootprintTemplates
Degreesof ShapeDissimilarity
Overlapped Footprint Shape
Person Recognition Method via Overlapped Footprint Shape
Dynamic Footprint-based User Identification System-39/59
Overlapped Footprint ShapesOverlapped Footprint Shape
Dynamic Footprint-based User Identification System-40/59
[Procedures for extracting aligned footprint]
When a dynamic footprint is given,Step 1. Labeling of each blobStep 2. Finding COA points of each blobStep 3. Determine first foot / second footStep 4. Updating overlapped footprint imageStep 5. Creation of overlapped footprint imageStep 6. Finding the rotation angle using the principal axes of
overlapped footprint imageStep 7. Creation of aligned footprint,
Overlapped Footprint Shape
),(* yxF
),(*
yxF
Person Recognition Method via Overlapped Footprint Shape
Dynamic Footprint-based User Identification System-41/59
[User Identification Method]
When a dynamic footprint and ,i=1,..,U are given, (Num. of User = U)
[Type I: Dissimilarity]User ID =
Deg. of Dissimilarity =
[Type II. Similarity]User ID = Deg. of Dissimilarity =
Overlapped Footprint Shape
),(*
yxF ),( yxFi
∑
−
yx
i
iyxFyxF
,
2*),(),(minarg
)()(
)(maxarg*
*
i
i
i FPFP
FFP ∩
∑
−
yx
iyxFyxF
,
2**),(),(
)()(
)(**
**
i
i
FPFP
FFP ∩
*i
*i
where P(A) is total number of pixels which has bigger value than threshold value
Considering the foot size
Considering the normalization
Person Recognition Method via Overlapped Footprint Shape
Dynamic Footprint-based User Identification System-42/59
0.90919.09090.86368.6364average
0.501011
0.5150.51510
1.5152159
0.5100.558
1101.557
0.5101106
0.55055
1.5152204
250.553
1150.5102
0.50051
FAR(%)FRR(%)FAR(%)FRR(%)
Method for normalization
Nakajima methoduser
Similarity value =
where P(A) is total number of pixels which hasbigger value than threshold value
P ( A B ) P ( A ) P ( B )
∩
Given two images A and B,
∑ −yx
BA yxIyxI,
2)},(),({Dissimilarity value =
Nakajima method (Type I)
Method for normalization (Type II)
Overlapped Footprint Shape
Person Recognition Method via Overlapped Footprint Shape
Dynamic Footprint-based User Identification System-43/59
Discussion on dissimilarity value(Type I)
1 2 3 4 5 6 7 8 9 10 11
0
2
4
6
8
10
12
14
16
18
20
user ID
diss
imila
rity
(pix
els)
Result with the data of user ID = 1
1 2 3 4 5 6 7 8 9 10 11
0
2
4
6
8
10
12
14
16
18
20
user ID
diss
imila
rity
(pix
els)
Result with the data of user ID = 2
Changing range: 3 pixels Changing range: 4 pixels
Sensitive to noise
USER 1 USER 2
Overlapped Footprint Shape
Dynamic Footprint-based User Identification System-44/59
Feature Extraction
WalkingFootprint
Quantized COPT
Extraction
User Identification
First Foot’sCOPT Matching
UserID
RegisteredHMMs
Second Foot’sCOPT Matching
Degreesof COPTSimilarity
weightedsum
COPT: COP TrajectoryCenter-Of-Pressure Trajectory
Person Recognition Method viaCOP Trajectory
Dynamic Footprint-based User Identification System-45/59
COP trajectories (COPTs)Center-Of-Pressure Trajectory
Dynamic Footprint-based User Identification System-46/59
[Procedures for extracting aligned footprint]Step 1. Labeling of each blobStep 2. Finding COA points of each blobStep 3. Determine first foot / second footStep 4. Check whether current time is the ending time of first foot
or the start time of second footStep 5. Updating overlapped footprint imageStep 6. Updating COP trajectory of each footStep 7. Creation of overlapped footprint imageStep 8. Finding the rotation angle using the principal axes of
overlapped footprint imageStep 9. Creation of aligned COP trajectory
Center-Of-Pressure Trajectory
Person Recognition Method viaCOP Trajectory
Dynamic Footprint-based User Identification System-47/59
Characteristics of COP trajectory
• Varying length time-series data with noise (by quantization)
• Different COP trajectories between Left and right feet
• Different curvatures as users
Hidden Markov Model
Separate model for each foot
Direction-based quantization(Relative position-based quantization)
Center-Of-Pressure Trajectory
Dynamic Footprint-based User Identification System-48/59
Period of second foot COP trajectory
Period of first foot COP trajectory
COPtrajectory
Relative position-basedquantization
HMM-basedfootprint recognizer
Center-Of-Pressure Trajectory
[User Identification Method]
Person Recognition Method via COPT
Dynamic Footprint-based User Identification System-49/59
• Element of HMM– Set of hidden states: S={S1,S2,…,SN}
where N is the number of states in the model
– State transition probability: A={aij} for 1 ≤ i,j ≤ Nwhere aij=P[qt+1=Sj | qt=Si], 0 ≤ aij, ∑aij=1
– Set of observation symbols: V={v1,v2,…,vM}where M is the number of observation symbols per state
– Observation symbol probability: B={bj(k)} for 1≤j≤ N, 1≤k≤ Mwhere bj(k) = P[vk at t | qt=Si], 0 ≤ bj(k), ∑ bj(k)=1
– Initial state probability: ∏={∏i}, 1 ≤ i ≤ NWhere ∏i=P[q1=Si], 0 ≤ ∏i, ∑ ∏i =1
[1] X. Li, M. Parizeau, and R. Plamondon, “Training Hidden Markov Models with Multiple Observations-A Combinatorial Method”,IEEE Tr. On PAMI, Vol.22, No.4, Apr. 2000
Registration of HMMs: Baum-Welch Algorithm
Center-Of-Pressure Trajectory
[REF]
Dynamic Footprint-based User Identification System-50/59
∑∈
++
++
+
=
=
===
FSlT
ttjijt
ttjijt
ttt
ljObai
OpjObai
Ojsispji
)()()()(
)()()()(
),,(),(
11
11
1
αβα
λβα
λγ
∑
∑
∑∑
∑−
=
−
=−
=
−
= == 1
1
1
11
1
1
1
)(
),(
),(
),(ˆ T
tt
T
tt
T
tj t
T
tt
ij
i
ji
ji
jia
γ
γ
γ
γ)(ˆ 1 ii γπ =
ελλ ≤− )()ˆ( OpOp
∑∈
=
==
FSlT
tt
tt
liiOispi
)()()(
),()(
αβα
λγ
∑
∑
=
=∈= T
tt
vOtt
i
i
ikb kt
1)(
)()(ˆ
γ
γ
Registration of HMMs: Baum-Welch Algorithm
Center-Of-Pressure Trajectory
[REF]
[REF] C. M. Bishop, Neural Networks for Pattern Recognition, Oxford University Press Inc., 1995
),,( πλ BA=1) Initialization of HMMs : Equal prob. Dist.2) Calculation of prob. at each node (state)
3) Re-estimation of model parameters
4) If , stop. Else, continue to step 2.
∑=
=N
iT iOp
1)()( αλObservation evaluation
Dynamic Footprint-based User Identification System-51/59
User Identification Method
1( ) ( )T Tnew old oldw w wλ ε−= − Ζ Ζ + Ι Ζ
2 21 ( ) ( )2 old new old new oldE w w w w wε λ= + Ζ − + −%
Error function :
Weight-updating rule :
( )n
n thni
i
where is the error for n patternwε ε∂Ζ ≡
∂
λSmall : Newton methodLarge : Gradient descent method
Used : Adaptive updating with initial value 0.01
λλ
[REF] C. M. Bishop, Neural Networks for Pattern Recognition, Oxford University Press Inc., 1995
[REF]
Center-Of-Pressure Trajectory
iλ)(ii OpP λ=Given the prob. from U user’s HMMs ,i=1,…,2U (O: quantized COPT)
jj
SmaxargUser ID = , Deg. of Similarity = *j
LM(Levenberg-Marquart) Learning Method
)(i
iijj OpwS λ∑= *jS
Dynamic Footprint-based User Identification System-52/59
4.502.0550.0020.4520
6.502.5070.0025.0018
6.503.0970.0030.9116
5.503.0975.0030.9114
6.003.2755.0032.7312
9.504.3665.0043.6410
12.005.1480.0051.368
20.506.0085.0060.006
58.009.27100.092.734
23.509.32100.093.182
Max. (%)Average (%)Max. (%)Average (%)
FARFRR# of learning dataAs time goes on,
User identification via COP trajectoryCenter-Of-Pressure Trajectory
Dynamic Footprint-based User Identification System-53/59
Comparison of foot shape and COPT
1 2 3 4 5 6 7 8 9 10 11
0
2
4
6
8
10
12
14
16
18
20
user ID
diss
imila
rity
(pix
els)
Result with the data of user ID = 1
Min. FRR error = 8.64 %Changing range in dissimilarity = 3 pixels
(Sensitive to noise)
Min. FRR error = 20.45 %
Result by overlapped footprint shape Result by COP trajectory
Combination of two independent information
Information fusion
Dynamic Footprint-based User Identification System-54/59
Information Fusion
Feature Extraction User Identification
WalkingFootprint
OverlappedFootprint
ShapeExtraction
Quantized COPT
Extraction
Footprint ShapeMatching(Type-II)
RegisteredFootprintTemplates
Degreesof ShapeSimilarity
First Foot’sCOPT Matching
RegisteredHMMs
Second Foot’sCOPT Matching
Degreesof COPTSimilarity
weightedsum
COPT: COP Trajectory
UserID
weightedSum
RegisteredWeight
RegisteredWeight
Information fusion
Dynamic Footprint-based User Identification System-55/59
Unified algorithm for the extraction ofaligned overlapped foot shape and COP trajectory
[Procedures for extracting aligned footprint]Step 1. Labeling of each blobStep 2. Finding COA points of each blobStep 3. Determine first foot / second footStep 4. Check whether current time is the ending time of first foot
or the start time of second footStep 5. Updating overlapped footprint imageStep 6. Updating COP trajectory of each footStep 7. Creation of overlapped footprint imageStep 8. Finding the rotation angle using the principal axes of
overlapped footprint imageStep 9. Creation of aligned footprintStep 10. Creation of aligned COP trajectory Common step
Step for foot shapeStep for COP
trajectory
Information fusion
Dynamic Footprint-based User Identification System-56/59
Result of Information Fusion
0.141.36TOTAL
1.000.00USER 11
0.005.00USER 10
0.005.00USER 9
0.000.00USER 8
0.005.00USER 7
0.000.00USER 6
0.000.00USER 5
0.500.00USER 4
0.000.00USER 3
0.000.00USER 2
0.000.00USER 1
FAR (%)FRR (%)User ID
Result of the fusion of footprint shape info. & COP trajectory info.
20 40 60 80 100 120 140 160 180 200 2200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Each person’s test data
Out
put
Recognition result of a user
Static + Dynamic Footprints
Static Footprint
Dynamic Footprint
u1 u2 u3 u4 u5 u6 u7 u8 u9 u10u11
One sample from u1 (user 1)is compared with all other data.
COPT Foot ShapeFoot Shape + COPT
Information fusion
Dynamic Footprint-based User Identification System-57/59
Walking RailSystem
- Example of Personalized Service
- Person Recognition byOverlapped Footprint Shape
- Greeting/Messaging Service
Automatic Door-Opening Service SystemApplication Example
Dynamic Footprint-based User Identification System-58/59
Concluding Remark
1. Analysis of previous footprint-based user identification method
2. Performance enhancement of previous footprint-based user identification methodby using the overlapped footprint shape (from 86.55% to 91.36%)
3. Proposal of COP trajectory-based user identification (FRR error = 20.45%)
6. Proposal of multi-modal identification using both aligned overlapped foot shapeand aligned COP trajectory
8. Successful fusion of previous gait and footprint method (FRR error = 1.36%)
Dynamic Footprint-based User Identification System-59/59
Thank youfor your attention