Image Sequence Analysis: Segmentation and Tracking

Pierre Kornprobst


Abstract

This talk concerns analysis of low-resolution image sequences with a static camera. Two different analyses, based on different methodologies, will be presented.

1. Segmentation. Usually, motion segmentation and image restoration are considered separately. We propose to solve the two problems in a coupled way, allowing the motion segmentation part to positively influence the restoration and vice-versa. To this end, a theoretically justified problem is proposed, studied on the space of bounded variations. A suitable numerical scheme is then derived, using Gamma-convergence and Geman-Reynolds Theorem.

2. Tracking. Considering these regions, a higher level of interpretation is approached establishing how these regions are related to each other. This problem presents many difficulties: severe occlusions, merging/splitting objects and defects in the detection. The method is based on a spatio-temporal (2D+t) representation of the moving regions. It uses a perceptual grouping approach, the Tensor Voting methodology, to enforce smoothness in the space and time of the tracked regions. Many examples will illustrate the talk.

On-line references

http://iris.usc.edu/home/iris/kornprob/User.html (Demonstrations and online publications available)


Maintained by Alexandre R.J. FRANÇOIS