Research On Binocular Vision Stereo Matching Algorithm

Stereo vision is one of the research hotspots in computer vision, which can acquire thescene’s depth information flexibly under a variety of conditions. Stereo vision is used toobtain the three-dimensional information of the scene by triangulation measure through thecorrespondence in images recorded from slightly different perspectives. Since stereo visionis an ill-posed problem, with complex scenes block, sparse texture areas and noise factors,stereo vision is still a research topic to be further development and refinement.In this thesis, several typical stereo matching algorithms are researched, we improveda stereo matching algorithm based on image segmentation from the viewpoint of obtainingthe high-precision dense disparity map of a scene. The selection of segmentation algorithmand the global optimization function have been improved. First we select Mean Shiftsegmentation algorithm to obtain some reasonable segments according the matchingrequirements, then obtain the initial disparity map by a combined weight area matchingalgorithm, a series of disparity planes are constructed using a robust fitting approachaccording the data of reliable segments which have more correspondences. And we do nottaking into account the parallax plane fitted by the unreliable segments. Finally, weconstruct a global optimization energy function and use graph cut method to distributethese reliable disparity planes and smooth disparity of some adjacent segments.In his thesis, Matlab are used to achieve the proposed algorithm, we also compare theefficiency of the proposed algorithm and other typical stereo algorithm in the standard testplatform for evaluation and comparison. The experiments confirmed that the proposedalgorithm has a high matching accuracy, clearly reproduce the object boundary informationand acquire the disparity of some weak texture segments.