Tracking and segmentation of moving objects in dynamic line images

As a part of a cine-film understanding system, this paper presents a dynamic scene analyzer which separates moving objects from the background and analyzes their motion patterns in dynamic line images such as cartoon films. Since the objects move and rotate in a three-dimensional world, and since occlusion often occurs, the shapes, sizes, and structures of the images of moving objects change from frame to frame. Moreover, the background or stationary objects may also move in the images because of camera movements to track interesting objects. The task of the analyzer is to segment the scene into meaningful constituents, and to obtain a structural description of each object containing properties, spatial relations, and motion patterns. Finding correspondence of regions and their segments in a sequence of input frames by a new flexible template matching method, the analyzer tracks moving regions and their segments in the dynamic images. Next, a similarity test of segment movements detects the background movement, and classifies the regions into stationary and nonstationary ones. Each in the latter group is further labeled as partly occluded, false, or moving, by examining both motion patterns of its segments and temporal change of its structure.   Finally, the analyzer merges segments of each moving object into groups, having similar motion patterns, in order to obtain a meaningful partition corresponding to its components such as hands or legs.