Research On 3D Scene Reconstruction Based On Binocular Stereo Vision

Binocular stereo vision is a crucial part of computer vision, which is widely used in aerospace, industrial, medical, transportation and other fields. Also, the manner that binocular stereo vision directly stimulates human eyes to process scene is simple and reliable. Therefore, research on 3D scene reconstruction based on binocular stereo vision has certain theoretical and practical significance. It uses two cameras to shoot the same scene and acquire the 3D coordinates of spacial point by making the use of the parallax of correspondence points in two images and projection transformation. And then,3D scene is reconstructed. The dissertation focuses on four key issues in scene reconstruction, and the main contents are as follows:(1) The linear and nonlinear model of camera imaging are studied, and the comparison of camera calibration method is made. In order to improve the precision of camera calibration, the internal camera parameters are obtained by second monocular camera calibration method based on board, and the cameras position relation is obtained by binocular calibration. The experimental results show that this camera calibration precision is higher.(2) Aiming at the problem of camera calibration affects the image rectification and image distortion, a new rectification method for non-calibrated camera are researched. In this method, we first extract the feature points by using SIFT algorithm and adopt RANSAC algorithm to eliminate error matching point. After that, the combination of polar rectification criterion and Jacobian determinant is for building energy function which is minimized to obtain projection transformation matrix for rectification. Finally, we can make the image distortion and rectification error smaller, and provides a standard image for subsequent stereo correspondence.(3) A Introduction of the basic steps and stereo matching algorithms are stated. Obtaining precise results of dense disparity map and improving the phenomenon that local stereo matching algorithm accuracy is not high, and real-time performance of stereo matching algorithm is weaker, A stereo matching improved algorithm based on graph cut is proposed. The improved algorithm uses the local SSD matching for initial matching parallax, which is improved by consistency checking criteria. Again, we can combine initial matching parallax with ?-expansion improved algorithm to construct energy grid which is minimized to acquire the disparity. Finally, high precision of dense parallax figure is obtained and the running time of correspondence algorithm is shorten.(4) In the model of binocular stereo vision, the three significant steps, camera calibration, stereo rectification and stereo correspondence are completed, dense point reconstruction can be realized by three dimensional measurement principle, which combined with the theory of spatial straight line reconstruction and Texture mapping technology to complement the The ultimate 3D reconstruction.