Dynamic SLAM Based on Neural Network and Depth Information

This is the authors' abstract. We don't add key points for this paper.

In the current Simultaneous Localization and Mapping (SLAM) framework, it is assumed that the environment remains static, meaning no changes in objects are detected. However, in real-world scenarios, encountering a dynamic environment where objects exhibit movement is inevitable. To address this issue, and we incorporate real-time object detection network You Only Look Once (YOLO) into our system to detect dynamic objects and mitigate their impact. Our approach utilizes an RGBD camera and develops an algorithm based on depth information to effectively distinguish between background and dynamic objects. By integrating an object detection thread into ORB-SLAM3, our system identifies dynamic objects within the environment while proposing a novel depth weight selection algorithm to maximize the number of static feature points for improved data associations. Finally, we evaluate tracking accuracy using publicly available TUM and BONN datasets.