Research On Local Optical Flow Estimation Methods

Optical flow estimation plays an important role in motion picture analysis. Most optical flow estimation algorithms can't get a good result in local regions. In this paper,though the research on textureless regions(untext) and discontinuity regions(disc),new method is designed to improve the accuracy of optical flow estimation in these regions. In order to get better motion results,many algorithms are used like block-matching. In addition,new neighborhood is found to calculate optical flow in discontinuity regions.First,many basic optical flow estimation methods are studied. Though the simulation images,it's clear to find that:Horn-Schunck method which uses regularization has bad results in the edge,and Lucas-Kanade method which uses local idea and least-squares algorithm has bad results in textureless regions.Second,new method is designed to reduce the error in these two regions. Though the research,most algorithms can't estimate the optical flow of all pixels. The pixels which can't be calculated are in the textureless and discontinuity regions. According to these findings,pyramid algorithm and block-matching method are used in textureless regions,assuming that every pixel moves parallely. As those pixels in discontinuity regions,for pixels in these regions always have different motions in most cases(sometimes multi-motions),so in such special regions,canny method is used to get edge images. Next,pixels which have the same movement with the central pixels are found to make new neighborhood.With these new neighborhood better optical flow estimation of the central pixels is got.Finally,these new methods are tested using the softwares of matlab and Microsoft Visual Studio 2008. Though the tests,new methods can slove the problem and get a more comprehensive and accurate result. In addition,absolute error and angle error are reduced.