Research On Image Matching Of Binocular Stereo Vision

Binocular stereo vision technology is a branch of computer vision, which is based on the parallax principle. Using different positions of the A, B or the same position a camera after moving the position or rotation shoots the same scene, calculates the corresponding points in two images of the position deviation and obtain 3D geometric information. Compared with other stereo methods, binocular stereo vision has the advantages of low cost, high efficiency, simple structure and appropriate precision. Since the establishment in 80 of the last century, it has become an important environment perception technology and is widely used in many fields of industrial detection, robot navigation, medical imaging, military affairs and aviation mapping.Implementation of binocular stereo vision technology in general can be divided into: image acquisition, camera calibration, feature extraction, stereo matching and 3D reconstruction. At the same time, stereo matching is the core and key problem and the current research focus and difficulty. Based on the analysis and synthesis of the results of the predecessors of the binocular stereo vision, stereo matching method is researched. First of all, the domestic and foreign research status of binocular stereo vision matching and the commonly used methods are described in detail and points out the advantage and disadvantage of feature matching, phase matching and region matching algorithm. Secondly, in view of the commonly used normalized cross correlation(NCC) region matching algorithm is existed large amount of calculation, real-time difference, matching the problem of low efficiency. By introducing the convolution theorem and integral image theory puts forward a kind of improved NCC algorithm and verifies the contrast analysis using the standard image and real image. The result of experiment shows that the improved algorithm increases the matching accuracy and speed. Third, the proposed algorithm is applied to the robot location. Through, the depth of information selecting different points of the results shows that the actual error is less then 1% and has a good effect. And the effectiveness of the algorithm is verified. Finally, this thesis is summarized and the future development is prospected.