Model Based Motion Field Estimation

Introduction The estimation of motion and boundaries of objects in an image sequence is an important issue for efficient video compression. It allows exploitation of the strong statistical bindings of image intensities along the motion trajectories [l, 41. The idea of object-oriented motion estimation is to subdivide the scene into regions of continuous motion. Thus discontinuities in the motion field may only occur at region boundaries. Ideally, each region uniquely corresponds to one surface of a moving object in the 3-D real world. The motion field is regarded as a pair of random fields V = (U, L ) , where U denotes a field of one motion vector per pixel, and L denotes a generic segmentation field. The segmentation field groups the motion vectors into several continuous regions.