Tracking Objects in Video Using Constellation of Parts

  • May 2006
  • Computer Vision

Abstract

Probabilistic modeling is a robust way to classify and recognize objects of some known category within an image. Constellation of parts is a relatively new method that achieves good results in recognizing objects in still images, and its model is capable of being built via unsupervised learning. It is proposed that the constellation of parts method has the potential to be a useful basis for tracking modeled objects in video. An object tracking system is described which uses a constellation of parts implementation combined with temporal coherence methods. It is proposed that this approach can result in a robust tracker with greater efficiency than a naïve frame-by-frame application of the constellation of parts method. Results are discussed which show some successes for a partial implementation of the system.

Downloads

A frame from the first video of results, showing a car wih a bouding box drawn around it

DivX is a video codec. There are many ways to play DivX videos, including standalone commercial software. On all computer platforms, my preferred software for playing DivX videos is VLC, an open-source multimedia player. On Windows, you can create a DivX video by using the open-source VrtualDub.

Selected references

R. Fergus, P. Perona, and A. Zisserman, “Object Class Recognition by Unsupervised Scale-Invariant learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, vol. 2. June 2003, pp. 264–271. Overview, data, and links are available online.