Dense stereo matching based on edge constraint and variable windows

We present a novel method on dense stereo matching; both high accuracy results and a handling of occlusions can be achieved with the edge constraint we prove in this paper. Though a lot of efforts had been made to solve the problems such as occlusions and disparity discontinuities, dense stereo matching is still very challenging in the field of stereo vision. Variable window methods seem to be a good solution, but the existing algorithms usually have complex estimation of an optimal window for each pixel. In our algorithm, the edges detected in the stereo image pair are selected and categorized into two classes to form an efficient constraint, namely edge constraint. By using the edge constraint, the optimal windows for each specific match can be determined and pixels within the windows would be matched with propagation. Compared to the existing window-based algorithms, our method not only can handle the problems caused by occlusions in the stereo image pair, but also can achieve the dense disparity map accurately. Experimental results show the good performance of our method.