Sie sind auf Seite 1von 29

Gait recognition under nonstandard circumstances

Kjetil Holien

Disposition
Research questions Introduction Gait as a biometric feature Analysis Experiment setup Results Conclusion Questions
1/27

Research questions
Main research questions:
To what extent is it possible to recognize a person under different circumstances? Do the different circumstances have any common features?

Sub research question:


Do people walk in the same way given the same circumstances?

2/27

Introduction
Authentication can occur in three ways:
Something you know, password or PIN code. Something you has, key or smartcard. Something you are, biometrics.

Biometrics are divided into:


Physiological: properties that normally do not change, ngerprints and iris. Behavioral: properties that are learned, such as signature and gait.

3/27

Gait as a biometric feature


Three main approaches:
Machine Vision based. Floor Sensor based. Wearable Sensor based (our approach).

4/27

Machine Vision
Obtained from the distance Image/video processing

Unobtrusive
Surveillance and forensic applications

5/27

Floor Sensor
Sensors on the floor Ground reaction forces/ heel-to-toe ratio

Unobtrusive
Identification

6/27

Wearable sensors
Sensor attached to the body Measure acceleration

Signal processing
Unobtrusive Authentication

7/27

Performances of related work


Body location Ankle Arm Hip (our approach) EER, % ~5 ~ 10 ~ 13 Number of Subjects 21 30 100 50
8/27

Trousers pocket ~ 7.3

Gait analysis
Sensor records acceleration in three directions:
X (horizontal) Y (vertical) Z (lateral)

Average cycle method:


Detect cycles within a walk. A cycle consist of a doublestep (left+right). Average the detected cycles (e.g. mean, median). Compute distance between average cycles.
Euclidian, Manhattan, DTW, derivatitve

9/27

Average cycle method


Compute resultant vector: Time interpolation: every 1/100th sec Noise reduction: Weighted Moving Average Step detection Average cycle creation

10/27

Raw data, resultant vector

11/27

Time interpolation and noise reduction

12/27

Step detection (1/2)

13/27

Step detection (2/2)


Consist of several sub-phases:
Estimate cycle length Indicate amplitude details Detect starting location Detect rest of the steps

14/27

Creation of average cycle


Pre-processing methods:

Normalize to 100 samples Adjust acceleration Align maximum points Normalize amplitude Skip irregular cycles
Mean Median Trimmed Mean Dynamic Time Warping
15/27

Create average cycle:

Cycles overlaid

16/27

Average cycle, mean

17/27

Experiment setup
Main experiment:
60 participants, two sessions of collection. 1st session: 6 normal walks, 8 fast and 8 slow. 2nd session: 6 normal walks, 8 circle walks (4 left and 4 right).

Sub-experiment:
5 participants walking 40 sessions 2 months. Each session consisted of 4 walks in the morning and 4 walks in the evening.

Sensor was always at the left hip.


18/27

Results
Best results when:
Normalize to 100 samples. Adjust acceleration. Aligned maximum points. Removed irregular cycles. Mean and median average cycle. Dynamic Time Warping as distance metric.

19/27

Normal walking
EER, % Automatically Manually
1st session 2nd session All normal 1.64 1.94 5.91 0.66 1.04 4.02

20/27

Other circumstances
EER, %
Automatically Circle left 2.97 Manually 1.31

Circle right
Fast

5.96
3.23

0.90
2.94

Slow

10.71

4.80

21/27

All circumstances
Normal vs other circumstances
EER between 15-30%

Multi-template
1 template for each circumstance, the others as input EER = 5.05%

22/27

Common features
Cycle length:
Normal: [95..125], average of 109 samples Fast: [80..110], average of 96 samples Slow: [110..180], average of 137 samples Circle same as normal

Amplitudes related to cycle length

23/27

Long-term experiment (1/3)


Morning vs morning / evening vs evening
Compare sessions at different days intervals

24/27

Long-term experiment (2/3)


Linear regression to compute a linear function (y = a + bx). Use hypothesis testing:
H0: b = 0 (stable walk) H1: b > 0 (more unstable walk)

Results:
Rejected H0 for 90% distance increases as time passes by.

25/27

Long-term experiment (3/3)


Morning vs evening (same day) and evening vs the consecutive morning
No difference in the average scores. Between 30% and 70% increase compared with 1 day interval scores.

26/27

Conclusion
Extremely good EER when comparing the circumstance with itself. Different circumstances seems to be distinct hard to transform X to normal. Good results when using a multi-template solution. Gait seems to be unstable to some extent need a dynamic template.
27/27

Questions?

Thanks for listening!

Das könnte Ihnen auch gefallen