Research And System Implemention Of Tree-Dimensional Reconstruction Based On Binocular Stereo Matching

Binocular stereo vision is an important branch of Computer Vision, which is widely used in many fields, such as robot navigation, medical examination, military research, automated production and so on. Stereo match is a significant technology of stereo vision, it simulates the human visual system to make the disparity image from stereo images captured by binocular camera system for the3d geometric information, which is important for three-dimensional reconstruction.The integrated3d reconstruction system based on binocular stereo generally includes image capture, camera calibration, image pretreatment, stereo matching and3d reconstruction. This paper has done lots of research work for the stereo vision system especially on camera calibration, stereo matching and3D reconstruction. And we have implemented the3d reconstruction system.In the stage of camera calibration, we first have made a detail research on the basic theory of binocular stereo vision, including the principle of binocular stereo vision as well as the models for the linear camera and nonlinear camera. Secondly we have realized the binocular calibration by the plane viewing method proposed by Zhengyou Zhang, which makes the image pair matching the epipolar constraint in ideal binocular stereo vision system. In the stage of stereo matching, we have made a classification research on stereo matching on the basis of the steps for binocular stereo. And then we made an intensive study of stereo matching algorithms putting emphasis on the steps of matching cost computation and cost aggregation. By comparison with principles of algorithms and experimental analysis, we finally chosen a non-local cost aggregation stereo method based on the minimum spanning tree for our3d reconstruction system, which makes a good performance on the matching speed as well as robustness. In the stage of3d reconstruction, we present a method for the reconstruction of the object from3d space by the mapping relationship of the object and the image. Then we use OpenGL as a visualize tool to show the reconstruction results, and we also paste the texture wrapping by the reference image to improve the reality of the3d scene.At last we have implemented the system with Visual C++and proved the validity of the3d construction system with real scene.