Fast long-term motion estimation for H.264 using multiresolution search

In this paper, we propose an adaptive motion search window location algorithm which incorporates multiresolution motion estimation information and the motion vector information predicted from neighboring encoded blocks. To reduce the computational complexity, we employ the spatial and temporal search range reduction that we proposed in [1]. Our proposed fast search is implemented and tested in the emerging H.264 encoder reference software (JM5.0c, [2]). The proposed motion estimation algorithm is implemented in the low-complexity mode of motion estimation and mode decision ([3]). We compare the proposed algorithm with the full-search method which checks the distortion values of all the candidates in the given search area. With respect to the fullsearch method, our proposed algorithm has two advantages. First, due to the spatial and temporal search range reduction, our algorithm is faster. Second, due to the multiresolution search, our proposed algorithm can use larger motion search range with small additional complexity. The proposed algorithm can significantly reduce the computational complexity of motion estimation with slight increase in bitrate as compared to a full-search that covers the same motion range over which our proposed fast search operates. The proposed algorithm provides particularly significant gains for video sequences that contain large motions.