Image Matching Algorithm Based On Binocular Stereo Vision

Stereo vision is an important researching field in the computer vision, its purpose is to restore and reconstruct the three-dimensional scene based on two-dimensional information. Stereo vision has been widely used in aerospace, surveillance, human computer vision, medical imaging, mobile robotics, industrial inspection,3D measurement and 3D stereo film etc. Stereo matching is the core of stereo vision research, the result of the stereo matching has more influence on the result of the 3D reconstruction. Stereo matching is an ill-posed problem with the influence of distortions, uneven illumination and self-occlusion in the scene. In this thesis, the stereo vision principles and binocular stereo matching methods have been studied, then the improved algorithms was proposed based on the purpose of improving matching efficiency.This paper first analyzed the background, purpose, significance, and the present research situation of stereo vision at home and abroad.Secondly, for the problem of the traditional local algorithm is inexact in the weak local texture region, occlusion area and sometimes is mismatching, a stereo matching algorithm based on pixel classification rectification and optimization was proposed. Firstly, the initial disparity was generated using the local matching stereo model. And the pixels were classified adopting correlation confidence measurement and the detection of texture-less regions. Then their disparity values were rectified.The final disparity map was obtained form segment-based optimization the rectified disparity.Third, this paper analyzed the process of belief propagation algorithm, and the construction and optimization of the energy function were studied. Concerning the high computation complexity and low efficiency in traditional stereo matching method based on belief propagation, a fast converging stereo matching algorithm based on sum of absolute differences (SAD) algorithm and BP algorithm was proposed. Firstly, the SAD matching method was used as a similarity decision criterion to determine the initial disparity map. When the energy function was constructed, the initial parallax map was used as a limit of function to get the optimization of disparity distribution by BP. And when the confidence level of each pixel was calculated in the optimization process, the convergence was considered, then the nodes of graphs were reduced and the convergence speed was improved.Aiming at increase the matching efficiency, the local matching and global matching algorithm was studied and the algorithms were improved, then the experimental have demonstrated the feasibility and effectiveness of the improved algorithm.