Semantic SLAM with Accurate Point Cloud Map Based on Dynamic Feature Point Removal

In dynamic environments, camera localization accuracy will be interfered by dynamic feature points, resulting in lower accuracy of camera pose estimation. Furthermore, traces of information regarding dynamic objects would remain within the generated point cloud map. In this paper, a module for removing dynamic feature points is put forward. By integrating the proposed module into the visual front end and introducing the map construction module, we develop a visual SLAM that is capable of creating an accurate point cloud map and running stably even in dynamic environments. Due to the accurate pose estimation, the proposed SLAM system can obtain point cloud maps with correct spatial geometric structure, and eliminate the smear introduced by dynamic objects by removing the dynamic regions. The proposed SLAM system was verified on the TUM dataset. Compared with ORB-SLAM2, the average optimization rate of ATE and RPE in RMSE under high dynamic environments is 94.39% and 41.48%, respectively. In addition, thanks to the removal of dynamic feature points, our system's uptime was reduced by 4.1% relative to 0 RB-SLAM2. Experiment results show that our system alleviates the influence imposed by dynamic objects.