A Visual SLAM System in Dynamic Environments Based on ORB-SLAM3

Vision SLAM has remained a central problem in robotics in recent years. Traditional vision SLAM methods often assume a static environment, but dynamic environments are more common in real-world scenarios. In dynamic environments, static assumptions often pose many challenges, such as feature point tracking, map construction, and accuracy degradation in localization. This paper presents a visual SLAM system for dynamic environments, built on ORB-SLAM3. An additional target detection thread is introduced alongside the original three threads to identify dynamic objects, enhancing system accuracy. First, the improved YOLOv7 is used to detect and remove dynamic feature points in the scene in real time, and then integrates semantic information, depth data, and optical flow to detect dynamic regions and filter out dynamic feature points., which assists the system in localization and map building. We validate our algorithm on TUM dataset and Bonn dataset respectively, and the results demonstrate that our algorithm provides higher localization accuracy and greater robustness than the original ORB-SLAM3.