Predictive 3D search algorithm for multi-frame motion estimation

Multi-frame motion estimation introduced in recent video standards such as H.264/AVC, helps to improve the rate-distortion performance and hence the video quality. This, however, comes at the expense of having a much higher computational complexity. In multi-frame motion estimation, there exists strong temporal correlation between the reference frames, which is not efficiently exploited in single-frame block-matching algorithms. In this paper, we propose a 3D motion search scheme which exploits the temporal correlation by using new 3D search patterns and motion vector predictors to obtain more accurate search centers. Compared to full search, our proposed algorithm results in PSNR losses of within 0.2 dB, while achieving a significantly lower motion estimation time by at least 96%. Furthermore, our results show that the proposed scheme is also significantly better than existing fast motion estimation algorithms for high-motion sequences.