Real-time Binocular Stereo Vision Research

Binocular stereo vision technique, which directly simulate man's eyes processing information manners has good adaptability and agility. It becomes the most important distance finding technique of wave optics three-dimension sensing methods, and it can be widely used in many domains. These years with computer processing faster and image captured more quickly, it is possible to realize real time binocular vision system from hardware and algorithm two factors. Binocular currently is a hotspot appealing many boffins. Based on Marr's theory, binocular stereo vision's basic theory, method and related techniques have been researched in this paper, aiming at realizing fast and even real time system by improving algorithms, so that it can provide references for the setting up of hardware system. The key contents are composed of five parts: fundamental and algorithm theory analysis, video capture, camera calibration, image matching and software designs. We give prominence to camera calibration and stereo matching. In order to calibrate camera's parameters, we skillfully choose plane grid template as frame of reference. Through extracting their feature points by using SUSAN operator, and applying linear method integrated with Levenberg-Marquarat algorithm nonlinear optimization to calculating, we finally solving sequence of the parameters rationally and get nice results. As far as stereo matching is concerned, a dynamic programming basing on mean normalized area correlation has been proposed. This approach firstly searches matching points candidates according area correlation, and then introduces dynamic programming to global optimization. Correspondences are finally obtained through finding an optimal path on a 2D plane composed of cross correlation. During the matching, The Box Filtering, pyramid frame, order consistency constraint and disparity searching range limiting are used to increase the speed of matching process. Dense disparity map are efficiently and nicely extracted by apply this method to rectified images. In the end, this paper shows the constitution modules of a real time binocular stereo vision system. Using an object-oriented design method in VC++, it develops related algorithm software including image capture, feature extract, camera calibration and dynamic matching. The calculating results accord with expectation.