Research on Improved Census Binocular Stereo Matching Algorithm

Abstract

Aiming at the defects that the traditional Census algorithm uses a fixed window and a fixed threshold to cause the image to have discontinuous depths and low matching accuracy in weak texture regions, an improvement is proposed. The cost computation phase uses SAD-Census algorithm, and proposes a new type of adaptive window method. The gradient information is used to dynamically select the threshold value to realize the selection of the window, and the Census cost computation is optimized. Consider the whole picture, Complete cost aggregation at multiple scales based on minimum spanning tree(MST); introduce left and right consistency detection methods to detect mismatched points in occluded areas, smooth the image through singular point filling and median filtering, and improve the overall accuracy of the improved algorithm. Using Middlebury dataset for testing, the experimental results show that the improved algorithm proposed in this paper has significantly improved matching accuracy and robustness compared with traditional algorithms, especially in areas with deep discontinuities and weak textures.