Research On Fast Motion Estimation Algorithms Based On Search Experience

As the key technology of video compression, motion estimation is significant for removing the temporal redundancy and improving video compression efficiency. As the video with high resolution and high rate develops, video application has increasingly performance requirements to video compression technology. In order to improve the video compression efficiency, video coding standards such as H.264/AVC and H.265/HEVC introduce a variety of new technologies in motion estimation module, including variable block size matching, subpixel precision motion estimation vector and multiple reference frames prediction and so on. However, the computational complexity rises markedly as these new technologies improve the precision of motion estimation. To reduce the computational complexity of motion estimation process, it is necessary to study and improve the traditional motion estimation algorithms.Since the search direction and the search step length are both uncertain in the search process of integer pixel motion estimation, the optimal matching point can't be located quickly, which leads to that the point set that need to search is large. Therefore a fast integer pixel motion estimation algorithm based on search strategy prediction is proposed. The proposed algorithm makes full use of the search experience of the encoded neighborhood blocks, and the correlation of search experience between neighborhood blocks, so as to select the appropriate search direction and search step length adaptively in the period of current block's search and improve the search strategy in the existing motion estimation algorithms. In addition, an early termination strategy for static blocks and a dynamic search window method are presented according to the correlation of the residuals at the optimal points and the motion vectors between neighborhood blocks, respectively. In order to evaluate the efficiency of the proposed algorithm, comparison experiments on searching time and video coding performance are conducted using typical algorithm and the proposed algorithm. Experimental results show that the proposed algorithm is an efficient ME algorithm, which greatly reduces the average searching time in integer pixel motion module while maintaining the coding performance of the conventional algorithms.Targeting the poor robustness of the subpixel motion estimation algorithm based on parabolic model and the high computational complexity of the interpolation based subpixel motion estimation algorithm, quantitative and qualitative analysis is performed on the residual characteristics of inter frame prediction in subpixel domain. Combining the two algorithms above, this paper proposes a fast subpixel motion estimation algorithm using adaptive prediction scheme by the parabolic model. The proposed algorithm predicts the residual surface characteristics of the current block utilizing the information of searched integer points, the search experience of the encoded neighborhood blocks, and the correlation of search experience between neighborhood blocks, and then selects the most appropriate search algorithm for the current block adaptively. Experimental results revel that the proposed algorithm greatly reduces the average searching time in subpixel accuracy motion module while keeping the almost same video coding performance compared with conventional algorithms. The proposed method is an efficient and robust algorithm.