Abstract: Aiming at the problem that feature points in ORB-SLAM3 algorithm are easily lost and have low accuracy, which in turn leads to large errors in motion trajectory of binoculars in complex scenes, this paper designs an improved ORB-SLAM3 algorithm. Firstly, the adaptive corner point detection technology is introduced in the ORB feature matching algorithm to increase the number of feature point acquisition. Secondly, the optical flow method is used to track the image features to improve the success rate of key frame creation. Then the region search is done with the feature points as the center to improve the real-time performance, the bi-directional left-right consistency test is used to screen the optimal parallax, the Prosac algorithm is applied to remove the mis-matched point pairs. Finally, the depth information is combined with the key frame. the depth information is combined with the key frame screening to improve the quality of key frames and optimize the camera pose. The improved algorithm has good robustness and positioning accuracy in absolute trajectory error.

Key words: binocular vision, ORB-SLAM3, optical flow method, Prosac algorithm