Research On Occlusion Problem In Stereo Reconstruction

Binocular vision is a hot issue in the field of computer vision research. Two cameras are used to capture different perspective image of the scene from different angles in the binocular vision system, and then 3D information is restored through the stereo matching process. As computer technology develops, binocular vision is becoming more and more widely used,such as virtual reality, reverse engineering, medicine, archaeology, entertainment and so on.Stereo matching as a very important problem of binocular vision,which has important significance and value to be further studied. And occlusion problems as difficult problem in stereo matching, also received extensive attention of the researchers. The stereo matching and occlusion detection of binocular vision are studied in this paper.First, a stereo matching algorithm based on dynamic programming is chose and improved in this paper after analyzing of all kinds of a variety of matching algorithm. Stereo matching results accuracy depends on similarity measure function,a new AD with Census transform similarity measure function is designed in this paper. The proposed similarity measurement function can overcome the influence of light transformation and noise to generate high quality initialization disparity search image, and improve the stereo matching accuracy effectively avoid the "stripes" phenomenon in the traditional stereo matching based on dynamic programming algorithm.Secondly, a new global matching cost function is constructed, the smooth constraint is redesigned and a dynamic method to alter the search radius is designed in this paper. The two methods assist each other. To maintain the disparity smooth, the methods reduce the search radius in the relatively small image gradient area; to allows the disparity jump, the methods increase search radius in the gradient in the relatively great image gradient area. Thus guarantee the correctness disparity value of the disparity discontinuous area, and greatly reduces the disparity search time.Again,a kind of occlusion detection algorithm based on left-right consistency check is designed in this paper, which can effectively improve the detection accuracy of occlusion area.After detecting occlusion area correctly, an occlusion area disparity optimal interpolation algorithm is designed based on color and distance similarity. This algorithm can keep the contour of the tested object completely and clearly, whose result is more close to the ground truth than direct extrapolation interpolation method, effectively improves interpolation accuracy in occlusion area.Finally, the proposed method is implemented using OpenCV image processing function library in Visual Studio 2010IDE. The stereo pair images provided by Middlebury college are tested and the results are tested in Middlebury college online evaluation platform, and 3D reconstruction experiments are done. The test results show that the algorithm in this paper is effective and advanced compare to other stereo matching algorithm based on dynamic programming.