To address the shortcomings of the current local stereo matching in weak texture regions with low matching accuracy and over-dependence on the central pixel, an adaptive local stereo matching algorithm based on the improved Census transform is proposed. Firstly, the Census transform is improved by using adaptive support window according to the texture complexity of the central pixel domain, introducing Tanimoto coefficients combined with Hamming distance algorithm, and fusing the absolute value of color or luminance difference as the new initial matching cost calculation. The cost aggregation is performed by the cross-cross domain algorithm and the winner-take-all algorithm is used to calculate the parallax. The left-right consistency method, iterative voting, interpolation filling and sub-pixel refinement are used in the parallax optimization stage, and the improved adaptive median filtering is used as noise suppression for edge blurring to obtain the final parallax map. The experimental results show that the proposed algorithm has an average mismatch rate of 4.39% on the Middlebury dataset, which is a significant improvement over other improved Census transform algorithms, and is robust and adaptive in terms of noise immunity.