Research On Stereo Matching Method Based On Region Growing

Eyes, is the main route of human perception from circumstances. The computers not only acquire, recognize and get the information from three-dimensional scene by modeling the human vision systems, but also make relates analysis. Computer vision can work in a variety of dangerous situations instead of the human eyes. So researches on machine vision is of quite significance. Stereo matching is a focus and challenge in computer vision, and how to achieve stereo matching quickly with high-precision is the key point for many scholars in recent years. Because stereo matching algorithm based on region growing has many advantages, this paper adopts it to do appropriate stereo matching for images. Main work of this paper are as follow:1) Illumination variation has a certain influence on the pixel gray of image. The matching error of traditional Census Transform algorithm results from the over-reliance on the central pixel’s gray. So this paper puts forward an improved Census transform algorithm with high robust. Firstly, a sum of gray average and local contrast is applied to replace the center pixel’s gray value in transform window, which improve the noise resistance of transformation results and the difference between different blocks. Next, it introduces the Gaussian template into the Census transformation to enhance the weight of pixels closed to center pixel, which effectively reduced the impact of discontinuous area on the matching. Finally, it adds the simulation lighting to the images of Middlebury and employs the actual matching value which is offered by Middlebury to calculate the matching precision in this paper. Experimental results show that the algorithm has a higher matching accuracy and a stronger robustness than traditional algorithm.2) The paper has studied the mechanism and the expansion steps of region growing algorithm. Confronted with low texture image based on traditional region growing matching algorithm, the growing can’t be of wide range because the seed points’ number is not enough. So a matching algorithm based on a fusion of SIFT features and contour information is put forward. First of all, the SIFT feature points and the information of the contour are extracted. Then, it uses two steps of matching method to finish the job of matching contour points:the first step, the feature points and epipolar constraints guide coarse matching for contour points; the second step, under edge correlation constraint, precise matching for contour points is completed through bringing in informations of the feature vector. Finally, seed points consist of the high-precision contour points and SIFT feature points perform the following regional growth and other processes. Theoretical Analysis and experimental results demonstrate that compared with the traditional region growing algorithm, the presented gets more seeds, larger expansion area and a higher accuracy.