A fast hierarchical radiometric invariant stereo matching algorithm

Abstract

Stereo matching aims at finding the corresponding pixels between two images. Most methods in the literatures are based on pixel intensity comparison. When images are taken at different illumination conditions or using different sensors, it is very unlikely that the corresponding pixels will have the same intensity creating false correspondences if it is only based on intensity matching functions alone. In this paper, a novel hierarchical radiometric invariant method is proposed to solve this problem. Pixels on different disparity planes in a local window are assigned the same label as the central pixel to solve the boundary problem. A more robust mutual information similarity function is used to compare the similarity of windows. A hierarchical structure is proposed to reduce the computational burden caused by using mutual information similarity function. Experimental results demonstrate that both the visual quality and quantitative analysis of the proposed algorithm outperforms most commonly used algorithms in the literatures.