In order to improve the accuracy and robustness of visual SLAM (Simultaneous Localization and Mapping) systems in dynamic scenes, a visual SLAM algorithm based on optical flow and instance segmentation is proposed. Aiming at the inconsistency of optical flow direction between dynamic objects and static background, feature points in the dynamic region mask can be eliminated in the original tracking thread of ORB-SLAM2 in real time. We use the existing depth map and tracking thread pose estimation information to remove the optical flow related to camera motion and then cluster the optical flow amplitude generated by the dynamic object's own motion to achieve high-precision dynamic area mask detection. The dynamic landmarks in the local mapping thread are eliminated combined with epipolar geometric constraints. Finally, the test results on TUM and KITTI datasets show that in high dynamic scenes, compared with ORB-SLAM2, Detect-SLAM, and DS-SLAM, the accuracy of the proposed algorithm is improved by 97%, 64%, and 44% on average. Compared with DynaSLAM, the accuracy has an average increase of 20% in half of the high dynamic scenes, which verifies that the proposed algorithm improves the accuracy and robustness of the system in high dynamic scenes.