Research On High-precision Stereo Matching Of Binocular Vision

Stereo matching is one of the research hotspots in the current computer vision.With the development of computer technology,stereo vision will be widely used in the robot vision、autonomous vehicle navigation、industrial measurement、object recognition,virtual reality and so on. However, due to the occlusion,noise and other factors, there isn’t a general method for stereo matching,so,stereo matching is the most important and the most difficult step in this field.Therefore the research of stereo matching is valuable. In this thesis,we will focus on the high-precision binocular stereo matching method.Starting from the camera model, this thesis introduced stereo vision model and theories related to stereo matching, the existing stereo matching algorithms were classified, the steps of these methods were summarized, and the general evaluation criteria for the stereo matching were provided.After the analysis of the problem that traditional median filter applied to the disparity map refinement process, this thesis presented an improved median filter for disparity refinement method:First, this thesis used the left/right consistency test and color segmentation to detect the mismatch regions and disparity discontinuities regions. And then filter out these regions in the filter window in order to achieve a better estimate of disparity for current pixel. Experimental results showed that this method could obtain a more precise disparity map compared to conventional method.In order to reduce the time complexity of stereo matching algorithm according to the further research of stereo matching algorithm based on the graph cuts, firstly, we could get an initial disparity map based on adaptive weighted window and the local optimization method; and then took the initial disparity as a reference and reduced the size of network, then energy function was optimized using graph cuts algorithm. The results showed that the stereo matching algorithm based on adaptive weighted window and graph cuts could get more accurate dense disparity map and significantly reduce the matching time.