A Linear Prediction Based Fractional-Pixel Motion Estimation Algorithm

In modern video coding standards, for example H.264, fractional-pixel motion estimation (ME) is implemented. Many fast integer-pixel ME algorithms have been developed to reduce the computational complexity of integer-pixel ME. With these advancements, fractional-pixel ME becomes the new bottleneck in the implementation of video encoders. For example, the conventional hierarchical fractional-pixel search (HFPS) at quarter-pixel accuracy requires computing the distortions of at least 16 fractional-pixel positions. This computation is comparable to or even higher than advanced fast integer-pixel ME process. Fast fractional-pixel ME algorithms were therefore developed, in which initial search point is first predicted before applying fast refinement search. Parabolic error surface model and fractional-pixel motion vector information of neighboring blocks are commonly used for initial search point prediction but they have problems of unsolvable solution and uncorrelated motion vectors, respectively. To tackle these problems, a center-biased fast fractional-pixel ME algorithm using linear prediction based search with intrinsic center-biased characteristic is developed in this paper. Experimental results show that the proposed algorithm is fast and robust.