Fast sub-pixel precision variable block size motion estimation

Fast sub-pixel precision variable block size motion estimation is a key issue for real-time application of the H.264. Many fast schemes have been proposed for fast variable block size sub-pixel motion estimation. Most of them are based on the analysis of the current frame. In this paper, we propose a new way to classify macroblocks of the current frame making use of the statistical information of the classified block types and motion activities of the previous frame, and the motion activities of the current frame. This forms a new and comprehensive scheme for fast variable block size motion estimation with sub-pixel refinement. With this prior knowledge, we can make decision on early rejection of sub-partition modes and early termination to skip candidate checking points. The new algorithm is able to alleviate substantially the computational effort of sub-pixel motion estimation. Extensive experimental work has been done, results of which show that our approach gives a speedup of 5 times over that of the fast algorithm in JM and a similar peak signal-to-noise ratio (PSNR) as the full search.