Research Of Adaptive Stereo Matching Algorithm Based On Improved Seeds Extending

Stereo matching is a key link of using two dimensional images to three dimensional reconstruct. The purpose of the stereo matching is to get an accurate disparity map from a pair of images, but because occlusion problem, illumination problem, baseline shift problem are existed in the left and right images, the difficulty of extracting the disparity is increased during image processing.Currently adaptive weight method and its improved algorithm is one of the most commonly used algorithms. Its principle is that each pixel is assigned a fixed window; calculate the weight of each pixel with the center pixel among the window, to get the weight of each pixel with the center pixel in the entire sub-window, then combined initial distance pixels to calculate the matching cost. However, due to its using the fixed window method, there are some inherent shortcomings, such as low matching precision in the sparse texture region and disparity discontinuity region. To solve these problems, this paper puts forward a new stereo matching algorithm, combination an improved adaptive window, adaptive weights and the seed point spread to disparity optimization.First, the calculation of the weight between pixels, since the adaptive weight algorithm has high mismatch rate in the low texture region and disparity discontinuities region, then analyzed, the proposed algorithm improves the weight set portion, distance weight and improved color weights.Second, the choice of the window, the cross-segmentation method proposed by Zhang, when excessive horizontal edge line in the image, i.e., disparity discontinuities region is horizontal rectangular form, matching rate drops, thus presents a cross segmentation method expanded simultaneously in horizontal and vertical directions.Third, for the initial disparity map obtained, according to the left and right consistency and blocking constraint strategies, puts forward a new seed point spread disparity optimization method that distinguish consecutive points and isolated points, for different points with different approach. The results show that the algorithm can get the disparity map with low mismatching rate.