Abstract:
The segmentation of an image sequence into meaningful regions is an important early step in computer vision. We describe a motion segmentation algorithm that tracks contours through an image sequence using active contours. This approach eliminates the problems of feature correspondence and feature drop-outs encountered with other feature-based tracking techniques. It also allows accurate recovery of the positions of segment boundaries and features. We present a simple texture-based energy equation that makes the active contours more stable and addresses the problem of "sliding" along edges. The method is capable of processing a frame per second without specialised hardware. The algorithm performs well, although active contours can be "captured" by occluding objects. Results are provided for real image sequences.
 

Downloads:

• A gzipped postscript version of the paper is available from here.
• An HTML overview is available here.