Fast Motion Estimation Robust to Random Motions Based on a Distance Prediction

Abstract

For fast motion estimation, a gradient descent search is widely used due to its high efficiency. However, since it does not examine all possible candidates within a search area, it suffers from PSNR degradation for sequences having fast and/or random motions. To alleviate this problem, we propose a hybrid search scheme wherein a hierarchical search scheme is selectively combined with an existing gradient descent search. For the selective combination, we introduce a measure estimating the distance between the current search point and the optimal point. Since this measure greatly reduces the need to perform hierarchical searches, their computational burden is not noticeable in the overall motion estimation while their contribution to the PSNR improvement is considerable. Using the estimated distance, we can also noticeably improve the early termination performance in a local search. Experimental results show that the proposed algorithm outperforms the other popular fast motion estimation algorithms in terms of both PSNR and search speed, especially for sequences having fast or random motions.