Segmentation and tracking using region adjacency graphs, picture trees and prior information

Tracking of video objects in pixel space using block based methods or template matching suffers from boundary alignment problems, assumes knowledge of object location, and is severely effected by rotation and shear. This paper suggests a region domain based object tracking algorithm. Using region adjacency graphs it is proposed to perform the search in region space, using relational information rather than absolute value. Interframe inconsistencies in region space are overcome using a picture tree search technique.