ing at the problems of noise-sensitive, easy distortion and with high false matching ratio in the disparity discontinuity region and weak texture region of the existing local matching algorithm, a multi-scale stereo matching algorithm for improved Census transform and gradient fusion is proposed. The weighted average gray value of all the pixels in the support window is used as the reference value of the Census transform. The Census cost is weighted combined with the gradient cost normalized by the horizontal and vertical directions, and a stable cost is obtained when the noise margin is set. Therefore, the reliability of the single pixel matching cost is obtained. Under the multi-scale, the improved guided filtering algorithm is used to complete the aggregation of the matching cost. The disparity map is obtained by parallax extraction. The experimental results demonstrate that the average false matching ratio of standard stereo image pairs obtained by the proposed algorithm is 4.74% on the Middlebury testing benchmark, and the average false matching ratio of the 27 extended stereo image pairs is 8.67%. In the parallax discontinuity region and the weak texture region, the false matching ratio is further reduced by the proposed algorithm, and it shows better robustness for noise and light.