Aiming at the problem that the traditional Census algorithm is too dependent on the center pixel, which is susceptible to noise, and the AD-Census algorithm can not make full use of the advantages of different algorithms, this paper proposes an improved Census transformation and adaptive weight stereo matching algorithm. Firstly, the mean value of the Census transform window and the pixel information of the center point and neighborhood in four directions are used to automatically classify the close pixels into one class, which improves the robustness of the Census transform against noise. Secondly, the SAD algorithm and Sobel edge detection are introduced, and the weight of SAD and Census transform is determined according to the gradient information, which improves the adaptability of the algorithm in different regions. Finally, the final disparity map is obtained by the cost aggregation method of the cross-domain and subsequent optimization. The parallax maps of different images are verified on the Middlebury platform, and the average error of the proposed algorithm is 9.33%, which is 3.39% lower than the AD-Census algorithm. Compared with other algorithms, the algorithm has better matching accuracy in the parallax discontinuous region and repeated texture region, and better robustness against noise and light.