A quadratic prediction based fractional-pixel motion estimation algorithm for H.264

Abstract

Motion estimation (ME) and compensation is the most important technique in video coding. In H.264, the motion vector (MV) of a variable size block is determined by performing the ME search procedure on integer-pixel positions, followed by fractional-pixel refinement. For fast integer-pixel ME algorithms, on the average, ME can be done by examining less than 5 search positions. Considering the conventional fractional-pixel ME algorithm, 8, 16, and 24 fractional-pixel search positions are required to be examined for the best MV at 1/2-, 1/4-, and 1/8-pixel accuracy, respectively. That is, the computational complexity of fractional-pixel ME becomes comparable to that of fast integer-pixel ME. Therefore, to develop an efficient fractional-pixel ME algorithm is greatly desirable. In this study, a fast fractional-pixel ME algorithm is proposed. In this study, a “degenerate” quadratic function is used to precisely determine the “best” quantized predictive motion vector (PMV) at 1/4-pixel accuracy for a variable size block. Based on the partial probability distributions of the sum of absolute component differences between the best MV at 1/4-pixel accuracy determined by the conventional 2-stage full search ME search algorithm and the “best” quantized PMV determined by the proposed algorithm, the search range of local fraction-pixel ME can be well determined. If the best quantized PMV determined by the proposed algorithm and that determined by the center biased fractional-pixel search algorithm are identically (0, 0), then (0, 0) is directly determined as the MV at 1/4-pixel accuracy for the variable size block, without applying the small diamond search pattern (SDSP). Otherwise, the SDSP at 1/4-pixel accuracy is used to determine the final result and the SDSP will be applied at most three times. Based on the experimental results obtained in this study, the four ME performance measures of the proposed algorithm are better than that of four comparison approaches, with slight degradations in average PSNR and bit rate.