Abstract: This paper proposes a modified Census transform which is based on information of the neighborhood for stereo matching to improve the classic Census transform by reducing mismatching of depth discontinuity as well as the noise disturbance. This new Census transform utilizes?2 bit to represent the differences between the pixel and its neighborhood. The result image of the transform provides more details on depth discontinuity and minimizes the impact of noise on the quality of matching. A dense disparity map is then generated through paralleled adaptive cost aggregation, stereo correspondence, sub-pixel interpolation and interpolation of the?occluded regions. The evaluation of Middlebury stereo images shows that the proposed algorithm enhances the accuracy of matching with concise structure and low complexity with strong robustness.

Key words: stereo matching, Census transform, related information of neighborhood, cost aggregation, depth discontinuity

摘要： 针对传统Census变换在视差不连续区域和噪声干扰情况下误匹配率较高的情况，提出了一种利用邻域相关信息的改进Census变换立体匹配算法。根据邻域像素的相关信息，将传统的Census变换中像素与邻域像素的差异应用2位信息表示，使变换后的图像在视差不连续区域的信息表示更为丰富，同时减少噪声对匹配质量的影响。通过并行化自适应匹配代价聚合、亚像素插值、左右一致性约束、遮挡区插值，最终得到了稠密视差图。经Middlebury立体图片测试表明，该算法结构简单，复杂度低，具有较高的鲁棒性，有效地提高了匹配精度。

关键词: 立体匹配, Census变换, 邻域相关信息, 代价聚合, 视差不连续