Human Identification using Silhouette Gait Data

1
Human Identification using Silhouette Gait Data

• Rutgers University
• Chan-Su Lee

2
Problem of Gait Recognition

• Advantage of gait as human identification
• Difficult to disguise
• Observable in a distance

• Difficulty of gait recognition
• Existance of various source of variation
  viewpoint, clothing, walking surface, shoe type,
  etc.
• Spatio-temporal image sequence Huge data,
  variation in speed-gtdifficult to compare

3
Standard Embedding of Gait Cycle

• Dimensionality of gait cycle
• One dimensional manifold in 3D space
• Half cycle-gt2D space with cycle
• Standard embedding on circles

4
Bilinear Models for Gait

• Gait Style
• Time invariant personalized style of the gait
• Gait Content
• Variant factor depend on time and viewpoint,
  shoes, and so on
• Represented by different body pose

5
Gait recognition algorithm(I)

• Asymmetric Model
• Symmetric Model

6
Gait recognition algorithm (II)

• Adaptation to new situation
• Learn new factor by modifying content vector
• Find style factor using new content vector

7
Experiment Results

• Improvement by normalized gait
• 14 peoples
• 3 different factors

8
Demos
Original Gait Data(GAR)
Different Surface(CAR)
Silhouette Images(GAR)
Silhouette Images(CAR)
Filtered Silhouette Images(GAR)
Implicit Function Representation of Silhouette
Images(GAR)
Normalized Gait Image Sequence(GAR)
9
Others