3D Scene Reconstruction Based On Binocular Stereo Vision

Binocular stereo vision is an important branch of computer vision, three dimensional reconstruetion based on Binocular stereo vision has been a hot research direction.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. It has been widly used in many field, such as medical, military, the protection of cultural relics and so on. It has very high scientific value and practical significance.First, this paper introduces the development process of computer vision, Marr vision theory and practical applications. Then it introduces the system of binocular stereo vision and imaging model. The constraint of the matching process, the classification matching method is introduced in detail. This paper adopts the horizontal binocular model, using the Bumblebee2 binocular camera, which based on geometric principle to recover the 3D scene. In the aspect of calibration, the Zhangzhengyou calibration method, using black and white checkerboard on camera calibration, the internal and external parameters of the camera are obtained. In image matching, respectively study the dense matching algorithm and sparse matching algorithm. In the aspect of dense matching, comparative experiments were carried out for SSD, SAD, NCC algorithm, obtained the disparity map. The introduction of integralimage matching theory, and the removal of the blind, NCC(normalized cross correlation) algorithm is improved. To improve the accuracy of the improved,operation time, decreased significantly. In the aspect of sparse matching, I study SIFT and SURF algorithm, for each step was compared with detailed description. As the three-dimensional reconstruction image of real-time and precision requirements are relatively high, the matching in image reconstruction in thie step is used for matching SURF algorithm, using RANSAC algorithm to eliminate false matching, experimental verification of the simulation. In the last part of the 3D reconstruction, the matching results in the formation of 3D point cloud, then the texture mapping of point cloud,get the final results of three-dimensional reconstruction image.