Walking Gesture Recognition

1

• Walking Gesture Recognition
• using Multi-Layer Perceptron
• CS679 Term Project
• by HoJun Son
• Computer Science Department
• KAIST

2
Contents

• Motivation
• Data Acquisition
• Normalization
• Feature Extraction
• Results
• Discussion

3
Motivation

• Immersive virtual environment
• restricted movements
• difficult to navigate
• need a intuitive navigation method
• Navigation method
• the walks in fixed position
• four gestures
• forward walk
• backward walk
• left-side walk
• right-side walk

4
Data Acquisition

• Magnetic trackers
• sensing a position and orientation
• sensitive to electronic smog devices
• noisy data are acquired
• Acquisition
• two magnetic trackers attached to two heels
• data sampling from 6 persons
• 4 gestures
• 100 left-right points per same gestures

5
Data Acquisition
P1
P2
P3
P6
P4
P5
6
Normalization

• Getting the minimum, maximum points
• Normalizing them between -1.0 and 1.0

7
Feature Extraction

• Cycle of a gesture
• 20 sampled data
• Feature vector
• consists of 9 feature points
• y-positions of 9 feature points are uniformly
  distributed by 0.25 between -1.0 and 1.0
• eg) -1.0 -0.75 -0.5 -0.25 0.0 0.25 0.5
  0.75 1.0
• x and z-position of feature point
• use the x and z-position of nearest one
  within confidence distance in y-position
  among sampled data
• otherwise, x, y, z-position are same as the
  nearest feature point

Sampled data
1.0
0.5
0.0
-0.5
-1.0
8
Feature Extraction
9
Results

• Training multi-layer perceptron
• input 27 (feature vector), hidden 15, output
  4 (4 gestures)
• Recognition statistics
• Recognition rate
• 95.8
• Mis-recognition
• backward gestures are mistakenly recognized as
  forward gestures

10
Discussion

• Magnetic trackers having higher sampling rate
• to prevent missing the highest point
• Starting and ending points
• to catch the cycle of a gesture
• More training data and experiments
• to prove this method
• Backward gesture
• if it works improperly, then it will be replaced