A Local Stereo Matching Algorithm Based on the Combination of Multiple Similarity Measures

Aiming at the difficulties in choosing matching cost and support window in stereo matching, a local stereo matching algorithm based on the combination of multiple similarity measures is proposed. Firstly, a matching cost is constructed, in which the census transform of image, the WLD (Weber local descriptor) feature of image, the color information of image and the gradient information of image are combined. Secondly, the guided filter is used to aggregate matching cost. Finally, a disparity refinement algorithm based on confidence and weighted filtering is proposed to eliminate the disparity choosing ambiguity brought by WTA (winner take all) strategy and horizontal stripe brought by LRC (left-right consistency) check. The standard test images provided by Middlebury test platform are used to test the proposed algorithm, and the percentage of bad matching pixel is 5.30%. Comparing with some high-performance algorithms, such as FastBilateral algorithm, the proposed method can achieve a higher matching accuracy.