High Performance Mathematical Quarter-Pixel Motion Estimation with Novel Rate Distortion Metric for H.264/AVC

Abstract

Fractional Motion Estimation (FME), which refines integer motion vectors found during Integer Motion Estimation, is one of the most time and computation consuming parts of the H.264 encoder. Conventional FME performs the time and resource consuming sub-pixel interpolation and subsequent secondary search, for each integer motion vector. In this paper using a novel rate distortion cost metric for sub-pixel selection, we introduce an improved mathematical method which estimates SAD values at quarter-pixel positions with far less computation compared to the conventional approach. The proposed scheme reduces both computation time and memory access requirements without any significant quality degradation. Simulation results indicate that the proposed method improves both PSNR and bit rate by 1 dB and %18 respectively compared to the basic mathematical method for fast motion samples. In the case of slow motion samples, it increases PSNR up to 3 dB.