A dynamic visual SLAM method based on ORB-SLAM3 for intelligent mobile robots

Accurately detecting and removing dynamic targets is crucial for enhancing the precision of visual simultaneous localisation and mapping (SLAM) systems in complex environments. To achieve high-precision and robust visual SLAM in dynamic settings, we propose a novel method called Dynamic-objects Semantic Visual SLAM, which integrates ORB-SLAM3 with YOLOv8. First, YOLOv8 is employed to detect and segment dynamic objects in real-time, and the feature information of these objects is seamlessly integrated into the ORB-SLAM3 front-end. Sparse optical flow tracking is subsequently utilised to track dynamic objects across frames, while enhanced ulti-view geometry addresses potential incomplete object detection issues in semantic segmentation. Finally, highly dynamic objects are filtered out to generate accurate localised maps. The dataset Technische Universität München (TUM) is used for experimental evaluation, and the results show that the absolute pose error is reduced by 78.63% and the relative pose error by 81.38%, significantly improving the success rate of mapping.