Abstract: In view of the fact that it is difficult for existing stereo matching algorithms to obtain high matching accuracy from images with radiometric differences, this paper proposes a novel algorithm based on Census transform and modified adaptive windows. First, according to the image structure and color information, an arbitrary shaped adaptive window based on the cross skeleton is constructed. Then, a matching cost based on Hamming distance is determined by Census transform. A two-step accumulation is used to reduce the computation complexity. The matching cost is subsequently optimized by winner-takes-all to gain initial disparity. Finally, A novel disparity refinement method based on mean-shift is proposed which is able to deal with the unreliable initial estimates and obtain a highly accurate disparity map. Experiments demonstrate that, compared with the state-of-art local algorithms, the proposed algorithm produces comparable accuracy: in particular, it can handle radiometric differences which are not solved by the state-of-art algorithms. Therefore the algorithm can be applied to environments of UAV vision navigation.

Key words: stereo matching, Census transform, modified adaptive window, radiometirc difference, disparity refinement