Aiming at the problem that the traditional ORB algorithm fails to meet the high-precision localization requirements due to the high mis-matching rate in the binocular feature matching stage, a feature matching algorithm based on the improved binocular ORB-SLAM3 is proposed. The nearest neighbor matching algorithm (FLANN) is introduced in the feature point matching stage, and more accurate matching pairs are filtered out by setting the ratio threshold, and the adaptive weighted SAD-Census algorithm is introduced in the binocular ORB-SLAM3 three-dimensional matching, and the geometric distances between the cases are taken into account to recalculate the SAD values and merge them with the Census algorithm to improve the stability and accuracy of feature matching, while the adaptive weighted SAD-Census algorithm is introduced. At the same time, the adaptive SAD window sliding range is added to further expand the search distance, so as to filter out the correct matches to improve the accuracy of the system. Experiments are carried out in the EuRoC dataset and real indoor scenes, and the results show that compared with the pre-improved ORB-SLAM3 algorithm, the localization accuracy of the improved algorithm is improved by 23.32% in the dataset, and nearly 50% in the real environment, thus verifying the feasibility and effectiveness of the improved algorithm.