Research On Key Technologies Of Three-dimensional Reconstruction In Binocular Stereo Vision

As a key research emphasis in the field of machine vision for a long time, binocular stereo vision system aims to extract three-dimensional information through two image sensors in different point of sight at same scenario and to reconstruct the three-dimensional scene, which perceives the world by simulating the human visual system. Based on the stereo vision theories, this paper’s achievements involve camera calibration, stereo matching, disparity map refinement and three-dimensional reconstruction.First the thesis built a simple and feasible operation image acquisition system, which included two industrial cameras. Then we finished monocular and binocular camera calibration based on Zhang calibration method. Stereo matching is the key technology of stereo reconstruction. Combining regional labels and SIFT algorithm, this paper proposed an improved SIFT algorithm. By setting regional restrictions to pre-matching points, both of the accuracy and efficiency of stereo matching are improved, and the executing time is reduced.In order to solve the problem of low accuracy for the disparity calculated by traditional matching algorithms, this paper proposed a disparity map refinement algorithm based on image region segmentation and region mapping. The algorithm first determined the minimum grey value of disparity according to the focal length, which was calculated by camera calibration. Then we considered the points whose grey values were lower than the minimum grey value as error matching points. Next, a reasonable texture filling process was applied on these points. The algorithm mapped segmentation regions of the original left image into the disparity map, and fitted planes with planar fitting algorithm in every region. This algorithm not only can remove mismatching points and disparity jumping points effectively, but also can preserve the edge details of texture.Three-dimensional reconstruction is the ultimate goal of binocular stereo vision. Combining disparity map and calibration data, three-dimensional coordinates of the points in the scene were calculated. Then the stereo models of the scene were generated. Experimental results show that our reconstruction results are reliable, further prove the availability of the camera calibration and stereo matching algorithm.