Brief Summary: Target Tracking from a moving platform

1
Brief Summary Target Tracking from a moving
platform

• Jackie Brosamer

2
Overview

• We want to track objects with a moving platform,
  using a map as a reference
• Local Association link detected regions within a
  sliding window and generate tracklets
• Global Association link tracklets and maintain
  track IDs

3
Map-Enhanced Detection

• Use global map such as a satellite image as
  reference frame for moving platform instead of
  first frame
• Reduced accumulated error
• Makes coordinates more meaningful (dimensions,
  latitude/longitude)

4
Geo-Registration

• First, use homography between consecutive frames
• Second, refine homography between image and map

5
(No Transcript)
6
(No Transcript)
7
(No Transcript)
8
Moving Regions

• For stationary, image sequence modeled at pixel
  level
• For moving, we fist model motion and then
  estimate background
• Adopt sliding window method

9
Local Data Association

• Maximize posterior of platforms to create
  tracklets
• Based on temporal compatibility within one track
  and spatial compatibility between tracks

10
Formulation

• Noisy Data Observations
• Find cover over time
• Based on
• Spatial association
• Temporal Association

11
MCMC Data associations

• Use monte carlo simulation to partition tracks
• Determine extension/reduction, birth/death,
  split/move

12
(No Transcript)
13
Global Tracklets Distribution

• Looks at longer time span to properly association
  tracklets with identity (esp when longer
  occlusion etc)