Stereo disparity estimation using adaptive window and dynamic programming

Abstract

One of the most challenging problems in stereo vision is to find corresponding image points in stereo image pair. In this thesis, a robust stereo disparity estimation algorithm is proposed and its experimental results are presented. The proposed algorithm combines the area-based search and the dynamic programming search. The robust disparity values are determined in some area by the area-based search according to certain criteria. The other areas where disparity values are not determined by the window-based search are estimated by dynamic programming. The proposed algorithm consists of three subparts - segmentation of homogeneous region and background, adaptive window search, and dynamic programming search. Most of the stereo algorithms may fail to find exact disparity value of homogeneous regions. Therefore, our algorithm extracts non-textured homogeneous region whose variance is less than a certain threshold by ordered seed filling algorithm. Adequate disparity value for such region is assigned by interpolation using the reliable disparity of the region boundary. We use adaptive-size window for the disparity estimation. The window size is determined adaptively from local variance for every pixel in a picture. If the matched result by the window search is not credible, the pixel point is reserved to a dynamic programming for more reliable estimation. We form the disparity space image for the unreliable pixel points, and applied the dynamic programming algorithm to them. The optimal path is obtained by tracing the path that has the minimum cost, so that final disparity value is acquired. Experiment results for several stereo images show that the proposed algorithm outperforms the window or dynamic programming only methods. We construct the left image with the right image according to the disparity value. Then, we compare the reconstructed left image with the original left image, and calculate the PSNR (Peak Signal-to-Noise Ratio) results in most of the image...