| Binocular stereo vision is one of representative techniques of computer vision andwidely used in many industry fields, therefore it comes rapidly into the focus of study.Stereo matching is the most important and most difficult step in binocular stereo visionand the main task is recovering three-dimensional (3D) geometric structure of visualsurfaces from stereo-pair images captured by binocular camera system from a 3D scene.Furthermore, it is also the important basis for three-dimensional reconstruction.Recently, the research of stereo matching algorithms is focus on a certain problemand has poor adaptability. After analyzing the problems of existing algorithms, thispaper discusses the main technique and achieves an area-based self-adaptive weightbinocular stereo matching system. This algorithm is mainly divided into three steps.First of all, we use a general algorithm called plane-based calibration to obtain theintrinsic and extrisic parameters by nonlinear iterative. The second step is pretreatmentfirst, including the Mean shift color over segmentation, extracting Harris corners andprojective rectification. Then, we propose a new stereo matching algorithm which isbased on the theoretical framework of area-based stereo matching method duringgetting the initial disparities and obtain a good result. According to the image context,the relationship of color and Euclidean space, occluding, a local window-based adaptiveweight algorithm is proposed. In the last step, a robust plane fitting technology isapplied to obtain disparity planes that are considered to be a label set. Finally, thedisparity plane assignment is approximated using an improved LBP algorithm which isdesigned for pixel-based method.We have implemented the system with Visual C++ and OpenCV. Two standardMiddlebury pictures and two pictures captured in experiments are taken for each part ofthe system. At last, experiment achieved good results, which proves the validity of thealgorithm. |
PDF Full Text Request