Dynamic SLAM Algorithm Fusing Semantic Information and Geometric Constraints

摘要

Traditional visual simultaneous localization and mapping algorithms usually assume that the environment in which the robot is located is static. When dynamic objects appear in the background, the localization accuracy and robustness of the algorithm will be affected by moving targets. Aiming at this problem, an effective VSLAM algorithm that adapts to dynamic scenes is proposed. First, the YOLOv5s object detection algorithm is used to extract the semantic information of the pictures captured by the camera, and the image scene is divided into a static area and a potential dynamic area according to the semantic information. Then, the algorithm in this paper only uses the feature points in the static environment when calculating the initial pose, and then combines the epipolar constraints and the weighted dynamic probability to eliminate the dynamic points in the environment. Finally, the initial pose is optimized by calculating all the static points in the environment. Experiments on TUM's RGB dataset show that the proposed algorithm is significantly better than the ORB-SLAM2 system in terms of trajectory accuracy and improves the robustness of SLAM systems in dynamic environments.