Quadtree based adaptive lossy coding of motion vectors

Many well-known video coding schemes use block-matching based motion estimation with a fixed block size, and motion vectors are coded using lossless entropy coding. The disadvantages of this method are that: 1) the predetermined block size is independent of the scene and may not be optimal; 2) the rate for encoding the motion vectors is fixed but may not be the optimal rate allocation between motion vectors and motion compensated prediction error. We propose a scene adaptive motion estimation and coding scheme based on the quadtree structure with entropy-constrained vector quantization (ECVQ). For each block size, an entropy-constrained vector quantizer is built to provide good motion vectors over many rates. A quadtree structure is then constructed that finds the best block size to perform adaptive motion estimation for each area of the frame, and thus optimally allocates rates among the quantizers. Our simulation results have shown that this coding algorithm has superior rate-distortion performance over the fixed-block-size ECVQ method and the traditional full-search method with lossless coding.