Aiming at the problem that traditional Census transform is susceptible to noise and the matching accuracy is low in the matching cost calculation process, a four-state Census transform algorithm with noise tolerance is proposed. In the calculation of the matching cost, firstly, the improved Census matching cost is fused with the matching cost of the absolute difference of grayscale and gradient, and the corresponding cut-off threshold is added to improve the reliability of the initial matching cost volume. Then, the cost aggregation is performed through the guided image filtering, and the strategy of “Winner Takes All” is used to calculate the initial disparity value. Finally, the final disparity map is obtained by optimizing the initial disparity value through left-right consistency checking, disparity filling and weighted median filtering. Experimental results show that the noise robustness of the proposed algorithm is better than that of the traditional Census transform algorithm, and the overall error matching rate of the stereo matching algorithm is reduced to 5.59%.