UMD VDT: Integration of Detection and Tracking Methods for Multiple Human Tracking

1
UMD VDT Integration of Detection andTracking
Methods for Multiple Human Tracking

• Son Tran, Zhe Lin, David Harwood, and Larry Davis

2
System Overview
Background subtraction
Human Detection
Track Logic
Region Tracking
3
Overview

• Green background subtraction blue human
  detection red final bounding box

4
Adaptive Background Subtraction

• Locally linear model of color change in response
  to illumination change for accuracy response
• Updating based on global prediction index for
  fast response

5
Adaptive Background Subtraction
1 200 400 600
800 1000
Linear update model
Ours
6
Human Detection

• Hierarchical part-template matching is employed
  to detect and segment humans.
• Shape information is probabilistically combined
  with region information to provide a reliable set
  of human hypotheses.
• Cascaded fast optimization scheme is applied to
  initial hypotheses to determine optimal subset of
  humans.

7
Implementation
(a) Input video frame
(b) Foreground blobs from BGS
(c) Foot candidate region
(d) Canny edge map in augmented foreground region
8
Implementation
9
Results Munich dataset
10
Tracking

• Based on regional affine-invariant features, MSER
• Occupancy map representation instead of
  appearance modeling

• Tracking Occupancy map construction

11
Putting it all together
12
And another