Personal Identification using Gait Data
on Slipper-device with Accelerometer
2021 Asian CHI Symposium
M i y u F u j i , K e i o U n i v e r s i t y
K a h o K a t o , K e i o U n i v e r s i t y
C h e n g s h u o X i a , K e i o U n i v e r s i t y
Yu t a S u g i u r a , K e i o U n i v e r s i t y
• Personal identification in entry / exit checks of indoor
facilities (elderly housing with care, community centre,
etc.) is significant;
• Understand the facility usage
• Reducing the burden on staff and users
2
Background: People gathering
3
Background: Personal Identification
・Use face and appearance
for identification
・Actions for identification;
Look, touch, etc…
・ Use behavioural
characteristics to identify
・Identified in daily activities
Burden on the user is small
Physical biometrics Behavioural biometrics
Knowledge based Property based
Biometrics based
・Key,IC-Card…etc
・Risk of theft or loss
・ID,password…etc
・Risk of leakage and forgetting
• Wearable device measures walking movement in
the facility
• Identify individuals in consideration of privacy
→Incorporating a sensor in slippers
8
Methodology
 Daily used in indoor.
 Easy to wear.
→Do not invade users'
daily life
Slippers installed at the entrance of the facility
9
Experiment:System
Learning
phase
Predict
phase
Data
Walking Feature
SVM
Classifier
System overview
Person
Prediction
result
10
Experiment:Feature extraction
Window based
segment
Hamming window
applied
FFT
( FFT )
Cut out only half
feature quantity .
SVM classifier
Uniform the
segmented data
Workflow of classifier building
Cross-validation
11
Experiment:hardware
Device specifications
Mounted device and sensor arrangement and their direction (6-sensor)
Sensor 3-axis accelerometer
Wireless module Xbee
MCU Arduino Pro Mini
12
Evaluation Protocol
Experiment1:Validation of foot-based
indentation
• Personal identification with gait dataset
Experiment2:Single feet based
identification
• Single data used, and considered the
optimal sensor placement
• IMU(Inertial Measurement Unity)based walking
dataset[7] for identification
13
Experiment1:Overview
Participant 10(Male 5・Female 5)
Motion Walking
Sensor position Full body 17 places
Frame rate [fps] 60
Length[s] 90seconds
Number of point 5000
Samples 128
Window size 120
Dataset Overview
Sensor location
[7] C. Xia and Y. Sugiura, "Wearable Accelerometer Optimal Positions for Human Motion Recognition," 2020 IEEE 2nd Global Conference on Life Sciences and
Technologies (LifeTech), Kyoto, Japan, 2020, pp. 19-20
14
Result & Discussion
Sensor position and accuracy
• The closer to the leg, the
higher the accuracy is.
• Right feet 94.3 %,left feet
95.3 %,both foot 97.0 %
• Foot based personal
identification is possible
• Optimumal sensor position assessment using a
one-feet slipper device
15
Experiment2:Overview
Participant 5(Male 2・Female 3)
Motion
Walking
(Do not indicate the speed or
step)
Environment long flat corridor
Sensed data
6 accelerometers on the right
foot
(Same slipper on the left foot)
Frame rate[fps] 37.5
Length[s] 32
Data points 1200
Samples 128
Window size 100
Overview
Walking status
• Accuracy using all sensors (6 locations) is 95%
16
Result & Analysis:Accuracy from all
sensors
All six-sensor based confusion matrix[%]
Six-sensor
17
Result & Analysis : Accuracy for each
sensor
Sensor position and identification accuracy
Sensor position and
name
Toe
Inner
Front
(IF)
Outer
Front
(OF)
Inner
Back
(IB)
Outer
Back
(OB)
Heel
18
Result & Analysis:More sensors used
Sensor combination and accuracy
93.3 % 88.3 % 88.3 %
93.3 % 91.7 % 91.7 % 91.7 %
19
Result & Analysis : Frequency domain
Change in frequency used
• Low frequency components may be highly
dependent on walking speed
• Considering the high frequency components
• Calculating identification accuracy by continuously
reducing the frequency range used from the low frequency
side
Sensor placement
20
Result & Analysis : Frequency domain
. Comparison by the number of sensors of average identification accuracy
when the frequency range used is changed
→Combine the 3 sensors, better accuracy is
expected.
21
Limitation and Future work
• Only the person registered as a data set can be
identified.
• New users need to get data for learning
→Proposal of a method to register a person who does
not exist on the dataset
• Only for straight-line walking on a flat surface.
• Data is acquired even in a state other than walking .
• Cannot identify movements other than walking, such as
going up and down stairs .
→ Combination with motion identification
Thank you!
2021 Asian CHI Symposium

Personal Identification using Gait Data on Slipper-device with Accelerometer - Asian CHI 2021 Symposium

  • 1.
    Personal Identification usingGait Data on Slipper-device with Accelerometer 2021 Asian CHI Symposium M i y u F u j i , K e i o U n i v e r s i t y K a h o K a t o , K e i o U n i v e r s i t y C h e n g s h u o X i a , K e i o U n i v e r s i t y Yu t a S u g i u r a , K e i o U n i v e r s i t y
  • 2.
    • Personal identificationin entry / exit checks of indoor facilities (elderly housing with care, community centre, etc.) is significant; • Understand the facility usage • Reducing the burden on staff and users 2 Background: People gathering
  • 3.
    3 Background: Personal Identification ・Useface and appearance for identification ・Actions for identification; Look, touch, etc… ・ Use behavioural characteristics to identify ・Identified in daily activities Burden on the user is small Physical biometrics Behavioural biometrics Knowledge based Property based Biometrics based ・Key,IC-Card…etc ・Risk of theft or loss ・ID,password…etc ・Risk of leakage and forgetting
  • 4.
    • Wearable devicemeasures walking movement in the facility • Identify individuals in consideration of privacy →Incorporating a sensor in slippers 8 Methodology  Daily used in indoor.  Easy to wear. →Do not invade users' daily life Slippers installed at the entrance of the facility
  • 5.
  • 6.
    10 Experiment:Feature extraction Window based segment Hammingwindow applied FFT ( FFT ) Cut out only half feature quantity . SVM classifier Uniform the segmented data Workflow of classifier building Cross-validation
  • 7.
    11 Experiment:hardware Device specifications Mounted deviceand sensor arrangement and their direction (6-sensor) Sensor 3-axis accelerometer Wireless module Xbee MCU Arduino Pro Mini
  • 8.
    12 Evaluation Protocol Experiment1:Validation offoot-based indentation • Personal identification with gait dataset Experiment2:Single feet based identification • Single data used, and considered the optimal sensor placement
  • 9.
    • IMU(Inertial MeasurementUnity)based walking dataset[7] for identification 13 Experiment1:Overview Participant 10(Male 5・Female 5) Motion Walking Sensor position Full body 17 places Frame rate [fps] 60 Length[s] 90seconds Number of point 5000 Samples 128 Window size 120 Dataset Overview Sensor location [7] C. Xia and Y. Sugiura, "Wearable Accelerometer Optimal Positions for Human Motion Recognition," 2020 IEEE 2nd Global Conference on Life Sciences and Technologies (LifeTech), Kyoto, Japan, 2020, pp. 19-20
  • 10.
    14 Result & Discussion Sensorposition and accuracy • The closer to the leg, the higher the accuracy is. • Right feet 94.3 %,left feet 95.3 %,both foot 97.0 % • Foot based personal identification is possible
  • 11.
    • Optimumal sensorposition assessment using a one-feet slipper device 15 Experiment2:Overview Participant 5(Male 2・Female 3) Motion Walking (Do not indicate the speed or step) Environment long flat corridor Sensed data 6 accelerometers on the right foot (Same slipper on the left foot) Frame rate[fps] 37.5 Length[s] 32 Data points 1200 Samples 128 Window size 100 Overview Walking status
  • 12.
    • Accuracy usingall sensors (6 locations) is 95% 16 Result & Analysis:Accuracy from all sensors All six-sensor based confusion matrix[%] Six-sensor
  • 13.
    17 Result & Analysis: Accuracy for each sensor Sensor position and identification accuracy Sensor position and name Toe Inner Front (IF) Outer Front (OF) Inner Back (IB) Outer Back (OB) Heel
  • 14.
    18 Result & Analysis:Moresensors used Sensor combination and accuracy 93.3 % 88.3 % 88.3 % 93.3 % 91.7 % 91.7 % 91.7 %
  • 15.
    19 Result & Analysis: Frequency domain Change in frequency used • Low frequency components may be highly dependent on walking speed • Considering the high frequency components • Calculating identification accuracy by continuously reducing the frequency range used from the low frequency side Sensor placement
  • 16.
    20 Result & Analysis: Frequency domain . Comparison by the number of sensors of average identification accuracy when the frequency range used is changed →Combine the 3 sensors, better accuracy is expected.
  • 17.
    21 Limitation and Futurework • Only the person registered as a data set can be identified. • New users need to get data for learning →Proposal of a method to register a person who does not exist on the dataset • Only for straight-line walking on a flat surface. • Data is acquired even in a state other than walking . • Cannot identify movements other than walking, such as going up and down stairs . → Combination with motion identification
  • 18.
    Thank you! 2021 AsianCHI Symposium