Research On Motion Estimation Algorithm Using Elastic Model

With the continuous improvement of video acquisition devices’ resolution, people put forward higher and higher requirements on the video coding efficiency. Motion estimation is a fundamental video coding technology. And most of the video coding efficiency improvements owe to more outstanding motion estimation algorithms. Therefore, the motion estimation not only attracts extensive attentions of researchers, but also is applied widely in practice.First, on introducing the motion estimation principle, this study discusses the research advances of motion estimation based on block-wise translational model from five aspects,namely the candidate vector down-sampling based motion estimation, the pixel down-sampling based motion estimation, the low complexity matching matric based motion estimation, the low bit-depth pixel based motion estimation, and the mesh based motion estimation, respectively. The advantages and disadvantages are comprehensively compared as well.Second, this study discusses the disadvantage of block-wise translational model when representing complex local motions. Furthermore, this study introduces several motion models, such as affine motion model. By comparative analysis, a conclusion is drawn that the elastic motion model is an efficient model representing complex motion field.Third, the elastic motion estimation algorithm based on discrete cosine functions is presented. According to the sensitivity of conventional elastic motion model to initial point and iteration step size, this study uses a 2-bit-depth-pixel search method to predict the initial iteration point, and computes the iteration step size by the ratio of low-frequency energy of discrete cosine transform and the golden ratio. An elastic motion estimation method is further proposed based on modified Gaussian-Newton algorithm.Fourth, this study introduces the idea of Levenberg-Marquardt algorithm to avoid the repetitive computation of Hessian matrix and its inverse matrix of the reference frame in the solution of conventional Gaussian-Newton algorithm. The Hessian matrix of the reference frame is replaced with the one of current frame, so that the Hessian matrix and its inverse matrix are only computed once in each motion estimation. Moreover, a halfway-stop strategy is employed. A fast elastic motion estimation method is thus addressed based on the modified Levenberg-Marquardt algorithm, so as to reduce computational complexity of elastic motion estimation.Extensive experimental results verify the effectiveness of the proposed two algorithms.