Robust stereo matching using improved ZNCC combined SAD-LBP

Abstract

Stereo matching aims to obtain depth information of scene from captured images, which becomes an active research topic in the field of computer vision. Most stereo matching cost algorithms are based on a common assumption, which is the intensity or color value of corresponding pixels are same. However, in real-world applications, the colors of the objects observed in the recorded image data are affected by radiometric variations. In this paper, using a novel similarity measure we propose a robust stereo matching method, which has robust performance to noise, illumination condition changes, and exposure changes between left and right images. The proposed stereo matching cost combines improved zero mean normalized cross-correlation (ZNCC) model and the absolute difference of local binary pattern (LBP) of windows to get both the color and texture similarity of windows to be matched. Based on Middleburry data set, we verify the effectiveness of the proposed algorithm. Computed results show that the proposed algorithm is more robust to illumination changes and noise than related stereo matching algorithms.