3D Reconstruction Based On Binocular Vision

With the continuous development of image processing and computer technology, computer vision as a burgeoning interdisciplinary subject has made a rapid development on the facet of both theory and practical application. And in the field of computer vision, three-dimensional reconstruction is always a hot issue. A lot of researchers at home and abroad are engaged in the study of this issue. With the continuous development of three-dimensional reconstruction technology, it has been used in many fields, such as robot navigation, image supervision and analysis, agriculture, architecture, military affairs, culture heritage protection and so on.Among all the methods about 3D reconstruction, the method based on stereo vision just need natural light and don't need set special light by people. The cost of this method is lower and the applicability is stronger, so this method is considered as one of the most widely-used 3d reconstruction technology.First, this paper introduced the computer vision, stereo vision implementation process and the status of the research about three-dimensional reconstruction based on binocular stereo vision at home and abroad, which elicited the goal of this paper, that was to achieve three-dimensional scene reconstruction via the binocular stereo vision system. Because real-time capability and accuracy of three-dimensional reconstruction are both important, this paper focused on four key technologies, namely feature extraction, camera calibration, stereo matching and the algorithm of three-dimensional coordinates. At the same time, many corresponding experiments had been done for each key technology.In the part of feature extraction, this paper researched corner detection and edge detection which was the basis for the next work. The paper analysde and compared the advantage and disadvantage of SUSAN operator and Harris operator. Then the paper improved the corner detection algorithm which combined SUSAN with Harris to detect the corners of the chessboard template. Otherwise, the paper also researched Roberts operator, Sobel operator, Laplacian and Laplacian-Guass operator. Through the contrast, the paper choosed Canny edge operator to detect the edges of images.In the part of camera calibration, the paper researched projective geometry, set up model of camera, analysed DLT, Tsai and Zhang calibration methods. According to the binocular stereo vision system which was used here, the paper choosed Zhang calibration method and combined the knowledge of feature extraction to get the internal and external parameters of camera. This calibration process took full account of the radial distortion of the lens, so the results of camera calibration were more accurate and also could be used in a wider range.In the part of stereo matching, after considering the demands about accuracy, speed and dense disparity map, this paper made use of region-based stereo matching algorithm as a theoretical basis and proposed the improved linear growth algorithm and error energy minimized method. Meanwhile five matching constraint conditions and filtering the unreliable disparity estimate were both considered in the improved algorithm. This algorithm achieved the stereo matching between the left images and right ones.Finally, the paper researched the model about 3D reconstruction and the methods about how to get 3D coordinates. At last, the paper choosed the method which was based on the disparity distance measurement method. This method changed the disparity map into a depth map or 3D view which was just the purpose of this paper.The paper used MATLAB and Visual C++6.0 to compile the corresponding algorithm. The results of experiments showed that the methods used in this article are correct and feasible. The work of the paper provides a basis for the research and application of robot.