Research On Stereo Matching Of Computer Vision

Visual stereo matching is one of the fundamental and significant problems in the study of the computer vision and contactless measurements. This technique makes it possible to reproduce a three-dimensional stereo by getting the distance information through pixels. On the other hand, it is the most difficult problem to be solved completely. In this work, the theory and algorithms for the stereo matching in the binocular vision are studied in detail. According to the area-based correlation, appropriate algorithms are developed with combining all the structural information including the contour and the feature points. Experiments on the real images show that the algorithm in this work is applicable and effective.Matching is performed directly with the image intensity in the area-based correlation without counting the space and structure correlation between objects. And therefore, the results are very sensitive to position, lighting, texture, and aberration. The disparity distribution function is determined by using the edge points as the initial matching set and hierarchical gaussian basis functions. Training error of feature points and disparities smooth constraint of non-feature points is included in the error function. Global optimum disparity function is estimated by gradient descent method.By using the high reliability of the edge feature points, this work develops the edge traction stereo matching method, which further improves the Barnard algorithm. The feature point confines the valid disparity range of the non-feature points. In the iteration process for the values in the disparity space, we fully consider the disparity distribution in the matching point's neighbor area, the relatively reliable edge feature points, and the different distributions. The window shapes of the supporting area of the non-feature points change according to the different distribution of the feature points. As a result, it reduces the uncertainty in estimating the matching values in the disparity space. In addition, the uniqueness and the continuity constraint are reflected by combining the examination of the conflict at sight lines and the examination and marking the occlusion area with matching threshold.This work develops a fast algorithm for the need of the real-time matching. A sparse initial disparity image of the edge points is produced with LoG operator. Using...