Matching Algorithm Based On The Three-dimensional Color Image Segmentation

Stereo is one of the significant problems of computer vision. The main study objective of computer vision is how to realize human being's vision function by using computer. It is very significant to study computer vision, because it is not only beneficial to satisfy the application requirements of artificial intelligence, but also to be helpful to deeply understand and study the mechanism of human being's vision system.With the development of the computer vision technology, the application of stereo vision is widely spreaded. Especially, with the merits of simple organization and easy-operating, the binocular vision is successfully used in the domains of industry examination, object recognition, work piece localization, robot self-leading, spaceflight etc. In these research area, the most difficult one is the stereo matching, which is also called stereo correspondence, and it is also the primary technology in stereo vision system. Though a large number of algorithms for stereo correspondence have been developed, there are still lots of problems needed to be solved.Firstly, the research content, its relative aspects in the world and its basic theory are introducted in the dissertation. The background and the main contents of this dissertation are described briefly too.Secondly, The camera is calibrated. By comparing the study of the previous calibration algorithms, analysising their advantages and disadvantages, a simple and effective method is proposed in the paper. First, the plane template is viewed from unknown orientation to obtain some images of the template. Then the vertex of every square are located as feature points. Afterward choose a reference frame is choosed and the world coordinates of each feature points is extractd. Based on the feature points which examines before each image's plane view matrix can be calculated easily, finally the camera parameter can be determined. Our camera is calibrated using this method. The result indicated that, it has good reliability, high precision and can simply implemented. It can widely used in project for its easy operation. Thirdly, a new implementation of mean shift and immersion simulation is presented to segment the image. Firstly, mean shift algorithm is used to seek the density center in the Lab space. Then, the density center is marked as remarkable points and their reflection in the image domain is marked as the start of immersion simulation algorithm. Finally through immersion simulation which area the points belonged to can be determined. The algorithm takes the space information as well as the color information. The proposed algorithm well accords with people's perception. The experimental result indicated that, all the points in one segmentation have the same color value, the segmentation edge tallies the actual object edge, the result satisfies the request of the stereo matching.Fourthly, an area-based algorithm is presented, it takes the image segmentation area as the matching elements. The traditional area-based matching algorithm adopts regular window which trends to obtain over-smooth results. The method presented in this dissertation applies the image segmentation to stereo matching, therefore it can solve the problem mentioned before. Compared with the other algorithm, the method fully utilize the whole color information because of color image segmentation. A layered configuration is employed to improve its efficiency. For large occlusion area and texture less area, the concept of believability and neighborhood correlation factor is presented. An area with low believability trends to take the biggest parallax value of neighborhood correlation factors, then the total matching cost is presented to ensure its precision. Finally, a correlation windows schemes is designed to meet different matching requirements. The experimental result indicated that, although the implementation of the algorithm is slightly slow, the mistake rate is small. For texture less area, the effect of matching is good, so the algorithm is effective.Finally, conclusions are given with recommendation for future work.