Research On Stereo Matching In Binocular Vision And Its Implementation

With the rapid development of electronics and computer technology, stereo vision has become one of the hotspots in computer vision area and its application field has been expanding. Binocular stereo vision has become a widely used model with the advantages of simple equipment, low cost, less human intervention. Stereo matching is one of the key technology in binocular stereo and its accuracy directly affects the effect of three-dimensional reconstruction. How to solve the problem effectively has important influence on the development of binocular vision. In consequence, there is not only important theoretical significance but also important practical value in the research of stereo matching.So far, solutions emerge in endless for the technical problems of weak texture, occlusions and disparity discontinuities. But how to improve stereo correspondence efficiency without the loss of matching accuracy remains a challenging research direction. In this paper, we intensively study the core theory and excellent algorithms of stereo matching around the binocular stereo vision system framework. And finally a new way of matching algorithm based on the combination of local and global technology is adopted, which comprises the advantages of both.Traditional dynamic programming (DP) algorithm has most advantage in efficiency among global classical stereo approaches, suffering from typical streaking artifacts because it only considers horizontal constraint in scan-line direction. Recently, a variety of improved DP algorithms have been proposed to orient to tackle this issue. Enhanced DP-based algorithm can reduce streak effects well by employing vertical consistency constraint between the scanlines, but with ambiguous matching at object boundaries. Local stereo approaches used to compute the matching costs volume within a finite window, with the limitation of dealing with the large textureless areas and occlusion boundaries. This paper presents a dense stereo correspondence algorithm combining local linear filtering and improved DP algorithm, which makes good trade-off between accuracy and speed.The performance evaluations using the Middlebury benchmark datasets demonstrate that our method produces results comparable to those state-of-the-art algorithms but is much more efficient.