Inferring Tracklets for Multi-Object Tracking

Problem

• Recent algorithms solve the tracking problem hierarchically: first find short tracks (tracklets), then iteratively extend into longer tracks
• How to optimally determine tracklets?

Approach

• Consider object detections in a sliding window of frames (4-8 seconds)
• Build an association graph (DAG), or detection graph (representing all possible associations over a sliding window), for each detection in the first frame of the sliding window
• Label detections in this graph as valid (consistent with the initial detection) or invalid (inconsistent)
  • MAP inference in a Bayesian network
• Generate track(s) from the resulting sequence of valid detections

Results

• clifresult.avi

Data

• freeway01

  Images: freeway01.zip (76 frames, 1207x1229, 26.6 MiB)

• seq21

  Images: seq21.zip (63 frames, 695x995, 30.9 MiB)

  Tracker configuration: seq21_config.xml

  Metadata: seq21_meta.xml

• seq30

  Images: seq30.zip (101 frames, 1312x738, 66.7 MiB)

Code

• Executables for Linux/Windows:

  2012-05-20 release (recommended) Download

  2012-04-09 release Download

  2011-08-16 release Download

• Source code release is pending review by USC Stevens Institute

Publications

• J. Prokaj, X. Zhao, and G. Medioni. Tracking Many Vehicles in Wide Area Aerial Surveillance. In Workshop on Camera Networks and Wide Area Scene Analysis joint with IEEE CVPR, 2012.

  [PDF] [Slides]

• J. Prokaj, M. Duchaineau, and G. Medioni. Inferring Tracklets for Multi-Object Tracking. In Workshop of Aerial Video Processing joint with IEEE CVPR, 2011.

  [PDF] [DOI] [Slides]

---