A Modified Census Transform Based on the Neighborhood Information for Stereo Matching Algorithm

Census transform is a non-parametric local transform. Its weakness is that the results relied on the center pixel too much. This paper proposes a modified Census transform based on the neighborhood information for stereo matching. By improving the classic Census transform, the new technique utilizes more bits to represent the differences between the pixel and its neighborhood information. The result image of the modified Census transform has more detailed information at depth discontinuity. After stereo correspondence, sub-pixel interpolation and the disparity refinement, a better dense disparity map can be obtained. The experiments present that the proposed algorithm has simple mechanism and strong robustness. It can improve the accuracy of matching and is applicable to hardware systems.