Region-based motion estimation in image sequences

We propose a method to segment moving objects in image sequences, so that each can be represented by its boundary and a compact motion description. Assuming that a scene consists of a small number of moving objects, we process each pair of frames as follows: first, we recursively identify a small number of “movement classes”, each represented by two or move motion parameters. Second, we use a spatially-segmented version of the reference frame (i.e. the later of the two frames) to classify segments into movement classes. Third, we merge segments using various similarity heuristics. Experimental results of some standard test sequences have shown that the proposed algorithm results in good quality motion segmentation with a small number of motion vectors