Fast local stereo matching with effective matching cost and robust cost aggregation

Stereo matching is extensively investigated in computer vision. This article is dedicated to local stereo matching based on adaptive support region for higher accuracy and speed. Firstly, we propose a novel cross-based and diamond-shaped sparse census transform with improved robustness and fastness compared with traditional methods. Secondly, an efficient computation of cost volume is proposed to accommodate diverse images. Once again, we improve the establishment of adaptive support region and ameliorate exponential step cost aggregation at the same time. At last, we achieve final disparity map by winner take all and disparity refinement. Experiments on Middlebury benchmark demonstrate our algorithm's better performance compared with other local methods.