For low bit rate video codec, many coding algorithms using translational block-motion estimation and compensation have been developed.
Accurate motion estimation of moving object is the key process of the object oriented low-rate coding of video image. In this thesis, a hierarchical motion estimation method based in an approximate and progressive motion model is proposed for the motions of two-dimensional planar object and of head-and-shoulder type real image.
First, three kinds of approximate 2-dimensional motion models are analyzed and compared with respect to accuracy and robustness. The motion description parameters of each model are 6, 4, and 2, respectively. It is shown that the choice of model is to depend on object size, and the type of motion.
Second, a hierarchical motion estimation algorithm based on a progressive motion model is investigated. The algorithm has three hierarchies of different motion models. It starts with the 6-parameter motion model for whole object, and then locally and progressively adjusts with 4 and 2 parameter models. It is shown that the proposed algorithm estimates somewhat more natural and accurate motion vector field than the conventional hierarchical block matching algorithm(HBMA) in a test image sequence.
Finally, as an application, the proposed algorithm is applied for image prediction in object-oriented analysis-synthesis coder. It is shown that the proposed algorithm reconstructs more natural images than the conventional HBMA. Furthermore, the required number of bits for motion information is much smaller than other methods, by about 700 bits/frame in average.