Video object tracking

S.Pateux. Contact: S.Pateux

Goal

In this study we are interested in the tracking of video objects in a video sequence. Objects tracked could be either rigid or non-rigid.

Approach

In order to track object masks, an algorithm based on a backward tracking technique has been used. This kind of algorithm allows to track rigid or non-rigid objects.

The principle of backward tracking is to perform for each frame in the sequence a classification of the regions obtained by a spatial segmentation. To this purpose, each region is backward projected on the previous frame according to its motion. Region is then classified to the object it backward projects mostly on.

In our approach, spatial regions can be further split in order to deal with too coarse spatial segmentation. Using this additional splitting allows to reduce artifacts introduced by spatial segmentation and to obtain more stable segmentation.

More details about this algorithm are given in [Pat00].

Results

Below are presented results obtained for various video sequences. Object mask were specified manually on the first frame and further automaticaly tracked for the next frames.

Foreman sequence

Intial Frame (71)	Frame 77	Frame 88

Animation: mpeg

Carphone sequence

Initial Frame (171)	Frame 271	Frame 379

Animation: mpeg (900Ko)

Other sequences

news (mpeg)	akiyo (mpeg 400Ko)

References

[Pat00]

PATEUX S. "Tracking of video objects using a backward projection technique". Proceedings of VCIP'2000 (Visual Communication and Image Processing). Perth, Australia. June 2000.