Research And Implementation Of GPU-Based Fast Motion Estimation Algorithm For Video Coding

Motion estimation is one of the key technologies in video compression systems,which use the pixels in adjacent frame as reference for the prediction of the pixels incurrent frame, and determines the efficiency and speed of video compression. Motionestimation is also the most complex module in computation, so fast motion estimationalgorithm was always adopted to reduce the unnecessary search for decreasing thecomputation complexity, according to distributed feature of the video sequence andmotion vector.In this paper, the implementation of fast motion estimation on GPU parallelprocessor for video compression was studied. On one hand, the proposed solutionreduced the unnecessary search according to distributed feature of video sequence andmotion vector. On the other hand, an accurate parallel fast motion estimation algorithm-GPU-based Local Full Search (GLFS) was proposed in this paper, which took use of thegeneral purpose computation of GPU. The whole motion estimation process is dividedinto the full pixel precision search and1/4pixel precision search, and each step in theprocess of the search is adopting a local full search algorithm.The algorithm partitionsthe search field into background and motion region by the content nature of videosequence, and detects whether the current block in motion region by local full search.Then, a coarse search was made in motion region with low resolution local full searchto find the optimal motion vector. Finally, a fine search with one fourth pixel accuracywas made. Furthermore, a stop condition was used in GLFS and the search steps in eachstage was constant.Finally, a GPU-based video codec was implemented on Windows platform withCUDA. The experiment results show that the search accuracy of GLFS is very close tothat of FS algorithm under the same video quality. Furthermore, the time consumptionof GLFS is only one ninth of the FS.