基于自适应权重AD-Census变换的双目立体匹配

In order to achieve a matching cost which involves the local feature information of pixels, a stereo matching system with the adaptive weight AD-Census was presented. The system improved upright cross support region construction by imposing the threshold of the absolute difference (AD) in gray between the adjacent pixels. An exponential function was used to adaptively set a pair of appropriate weights AD and the census matching costs based on the shortest arm of each pixel. Because the shortest arm of each pixel could reflect the local feature of pixels, the weights adaptively set was related to the local information of pixels. The weight of AD was larger than the weight of the census matching if the pixel was in regions nearby image edges. The weight of the census matching was larger than the weight of AD if the pixel was in textureless region. The experimental Results on the Middlebury stereo benchmark v3 showed that the performance of the system with the adaptive weight AD-Census is better than that of the system with AD-Census. The average disparity errors of all pixels are reduced by 25 % and the average disparity errors of non-occlusion pixels are reduced by 20 %. Moreover the matching results of some image libraries including Adir image library showed that the stereo matching system with the adaptive weight AD-Census is especially superior for the texture-rich images and the images comprising repetitive textures.