Research On Stereo Matching Algorithm Of Binocular Vision

Binocular vision technology is one of the most important investigated topics in computer visionfield, which is aimed at endowing the computers with ability of gaining three-dimension(3D)information from a pair of two-dimension(2D) digital images. Presently, this technology has beenwidely used in robot navigation,3D measurement, virtual reality and so on. As a research hotspot,stereo matching is a key step of binocular vision technology and plays a pivotal role throughout thewhole process.This paper starts from introducing the binocular vision model, roundly analyzes its workingprinciple and stereo matching theory, and induces the existing matching algorithms. According to theactual requirement, a new stereo matching algorithm is presented here, that is suitable for parallelbinocular vision system. The main job of this paper is as follows:1) The paper analyses the binocular imaging model, epipolar geometry, triangle measurement’sprinciple and the main key technics. And it introduces the main researching contents, and makes asummary of the existing stereo mathcing algorithms based on different units.2) The matching method can also be classified into local and overall matching method, accordingto their optimizing method. In the paper, some typical local algorithm and overall algorithm areresearched in detail, such as Winner-take-all, Realtime, Dynamic Programming and SimulatedAnnealing. The local matching method is simple and easy to be realized but without high precision,and the overall matching method is opposite. To combine the advantage of both local and overallmethod, this paper present a new stereo matching method based on the image segmentation with theepipolar constraint.3) About the image segmentation, this paper analyzes traditional methods, and presents asegmenting method based on active counter. This method expands the1D constraint to2D, and getsthe object’s counter in the2D region to reduce the searching area. For introducing overall information,the located region is reliable for matching. Compared with segmentation based on thresh, their meritand defect are analyzed.4) About local registration, this paper introduces a feature template with neighborhood difference,and improves the caculating process to cut down the matching cost. Compared with the traditionalmethod, the new method is not only simple but also robust.Finally, the synthesis algorithm is validated for its feasibility with different kinds of image pairs. The experimental results show that the algorithm has high speed and good robustness, embodying theadvantages of the combination between local matching and global matching methods.