Research On Stereo Matching Method Based On SGM

Stereo vision,which has certain application value in virtual reality,unmanned driving,3D reconstruction and other fields,is an important part of computer vision.Stereo matching is one of the key part in stereo vision,whose main task is to obtain the correspondence of the same point in different perspectives.Owing to the influence of illumination,texture and occlusion,stereo matching is difficult,which limits the development of stereo vision technology.This thesis discusses three kinds of stereo matching algorithms and chooses the Semi-Global Matching(SGM)algorithm as the research object.The algorithm has some problems,such as view occlusion,error matching and easy to be interfered by noise.Based on the stereo matching process,this thesis studies the algorithm from three aspects: cost calculation,cost aggregation and disparity optimization,the main content is as follows:(1)Combing Census transformation with filtering algorithm to calculate the cost.Census transformation is a kind of non-parametric transformation,which is not sensitive to the change of illumination.Census transformation can reduce the influence of illumination on matching result.Compared with the mutual information method adopted by the SGM algorithm,the principle and implementation of Census transformation are simpler and more efficient.Meanwhile,a filtering algorithm is added on the basis of Census transform in this thesis to reduce the noise interference and make the matching result more accurate.(2)On the basis of cost aggregation from multiple paths,the new penalty parameter is added to adjust the original penalty parameters.The original penalty parameters adopt fixed numerical method,which affect the quality of the disparity.Therefore,according to the change trend of adjacent pixels in any path,a new penalty parameter is added,and the original penalty parameters are automatically adjusted to make the aggregation result more accurate.(3)Performing disparity optimization on the obtained disparity map.When cost calculation and cost aggregation are carried out,the algorithm cannot avoid the wrong disparity caused by occlusion or non-optimal calculation methods,which will lead to error matching.In this thesis,the method of consistency check is used to eliminate the error matching points and fill the holes in the boundary area or the occluded area.Last but not least,the method of curve fitting can be used to replace the whole pixel with subpixel to make the disparity map smoother and more exact.Finally,this thesis uses the stereo vision algorithm evaluation platform Middle Bury to analyze the performance of the algorithm.Compared with the SGM algorithm,the AD-Census algorithm and the PMSGM algorithm,the error matching rate of the algorithm proposed in this thesis reduces by 9.79%,5.37% and 3.03% respectively.Compared with the AD-Census algorithm and the PMSGM algorithm,the running speed increases by 5.28 s and 13.39 s,showing the effectiveness of the algorithm in accuracy and efficiency.