Camera Calibration And Stereo Matching Technology Research

The purpose of stereo vision is to extract three-dimensional information from several different position images of the same scene.Stereo vision can be divided into five steps:camera calibration,feature extraction,stereo matching,3D information recovery and post-processing.This paper focus on the computer stereo vision system,aiming at the camera calibration and stereo matching technology,to lay a foundation to extract the 3D information from the 2D image.In the part of camera calibration,a method of Zhengyou Zhang which is between the traditional calibration and the self-calibration to obtain the inner and outer parameters of the binocular camera is adopted in this paper.Meanwhile,in order to facilitate the subsequent image matching,the distortion correction and the polar line correction are carried out on binocular images in this paper.In the matching of feature points,this paper compares the performance of several commonly used feature extraction operators and selects the SIFT algorithm with better anti-noise and stability for feature extraction and matching.Meanwhile,the RANSAC algorithm is adopted to remove the mismatch point,which is able to improve the matching precision.In the part of dense matching,this paper firstly studies dense matching method based on region growing,and expounds the concrete realization process in detail.Meanwhile,a dense matching method based on spatial homography is proposed in this paper.This method can realize by continuously assume that triangular patches formed by three pairs of adjacent feature points in the current registration results satisfy the homography relationship,then use the cross-correlation method to verify the homography hypothesis.The triangular facets which meet the principle of verification are recorded as the registration patches.The triangular facets which do not meet the verification principle are subdivided and re-judged.By this method,more dense matching points are detected from the image,so that the distribution of the matching feature points is more uniform and the matching accuracy is higher.