Robust Static Feature Extraction in Dynamic Scenes: An Advanced RGBD Visual SLAM Framework Leveraging YOLOv9

Dynamic elements in dynamic environment are important factors that affect the localization accuracy of visual Simultaneous Localization and Mapping (vSLAM), and may even lead to tracking failures. To address this issue, we propose an enhanced dense visual SLAM method specifically designed for dynamic environments. The whole framework of our approach is based on ORB-SLAM2, while integrating the YOLOv9 network to detect dynamic objects. By transmitting the detection results of YOLOv9 to the SLAM module, dynamic feature elements are then removed from the dense map construction thread. Finally, the dense map that only contains static information is created. The experimental results show that the filtering of dynamic objects not only can reduce the average root mean square error (RMSE), but also help and then improve the robustness of the SLAM system in dynamic environment.