Recognition of Human Gait From Video

1
Recognition of Human Gait From Video

• Rong Zhang, C. Vogler, and D. Metaxas
• Computational Biomedicine Imaging and Modeling
  Center
• Rutgers University

2
Outline

• Motivation
• Distinguishing features
• Recognition process
• Silhouette extraction
• Human model initialization
• Extracting joint angles over image sequences
• Recognition
• Preliminary Results

3
Motivation

• The goal is to detect and identify humans by the
  way they walk.
• The walking pattern (gait) is unique enough to
  identify a person.
• Such capabilities will enhance
• Human identification.
• Abnormal behavior detection.

4
Gait Cycle
5
Distinguishing features

• Features that seem unique to each person
• Joint angle between the upper and lower legs
• Relationship between the knee joints and the feet
  over time
• Elevation of knee joint over the ankle (i.e.,
  vertical distance between knee and ankle) shows a
  distinctive temporal pattern

6
Elevation over ankle is distinctive
Transition from swing leg to stance leg is
noticeably different across different people
over time
7
Gait Recognition Procedure
8
Silhouette Extraction Result
Background
Image
After background subtraction
Final result
9
Human Model

• Human is modeled by five connected trapezoids.
• Each trapezoid (body part) is represented by

10
Human Model

• Each configuration of human body is represented
  by
• where , and c as the
  center of the body.

11
Human Model Initialization Result
12
2D Model-based Human tracking

• Previous methods
• Cardboard person model
• Scaled Prismatic Model
• Twist and exponential maps
• Condensation-Based
• Our approach
• Tracking via Gibbs sampling(probabilistic)
• Advantages
• Able to handle occlusion implicitly
• Has greater chance of avoiding local optima

13
Tracking Results
14
Tracking Results
15
Recognition

• Collect feature vector with
• Elevation over ankle
• Joint angles between upper and lower leg
• Use left-right hidden Markov models for
  recognition
• One HMM per person, trained on a minimum of 4-5
  full step cycles from that person

16
Recognition (continued)

• Use algorithm similar to isolated speech
  recognition to identify people
• Collect a step cycle from test subject
• For each HMM in the database, compute likelihood
  that it matches signal of this step cycle
• Select HMM with maximum likelihood
• Person corresponding to that HMM is identified
  subject

17
Experiment

• Two sequence sets are taken at two locations one
  in a parking lot, one in front of a building
• Each set contains 3 persons walking sequences

18
Preliminary Results

• Perfect recognition scores across 3 subjects
• 5-6 step cycles per subject collected from the
  computer vision algorithm
• Use of HMMs with knee elevation and joint angles
  as features holds promise
• More work is needed to identify other
  distinguishing features

19
Thank You!