Research On Binobody Image Extraction Algorithm Based On Feature Matching

Stereo matching is the key part of binocular stereo vision. It generally divided into 4 modules: the matching cost, the cost of polymerization, parallax calculation and optimization, parallax correction. The depth of the object value can be calculated by these 4 modules. But it will be affected by image noise, multi texture, occlusion and a series of factors in the implementation.At present, the representative algorithms of stereo matching were as follows: the image segmentation algorithm, BP belief propagation algorithm, linear growth matching algorithm, adaptive window, and adaptive weight algorithm and so on. In this paper, the image segmentation algorithm, BP belief propagation algorithm and linear growth matching algorithm were studied, Comprehensive analysis of the three algorithms, linear growth matching algorithm found the root nodes as the search strategy, at the same time the judgment method of the barrier point can be expanded. So the further research based on linear growth matching algorithm were put forward.In this paper, the main innovations were as the following several aspects:1、Windows were divided according to the relationship of match pixels and the pixels around and matching function was put forward to further constrain the growth of the root node. All those were to solve the serious problem of linear texture in the depth obtained by linear growth matching algorithm.2、Feature point matching algorithm was put forward to solve plaque problem which caused by the root node error localization during using the linear growth matching algorithm to deal with the images with a large area of low texture. Harris corner points were selected as the feature points, and the "Ghost point" hypothesis was put forward, combined with improved window matching cost function to solve the false matching problem, and improve the accuracy of angular point matching.3、According to the problem only could get a sparse depth map by the Harris corner matching algorithm, angular point matching algorithm was proposed to combine with linear growth matching algorithm, and the matched feature points was used as the root node. At the same time nonlocal cost aggregation algorithm based on minimum spanning tree was increased to select new root node to calculate the dense parallax images.Through the experiment, more accurate depth map could be obtained by this algorithm. It could solve the difficult stereo matching problem in the shade, low texture area, but the execution speed of this algorithm also need to be further improved.