Research on RGB-D Visual SLAM Algorithm in Dynamic Environments

Currently, most visual simultaneous localization and mapping (VSLAM) systems focus on static environments, while dynamic objects in real-world scenarios significantly degrade system accuracy and robustness. To address this issue, we propose an improved algorithm for indoor dynamic environments. Based on ORB-SLAM3, we introduce a YOLOv8-based dynamic object detection thread that performs instance segmentation and generates semantic information for image frames. In the tracking thread, we employ a sliding window mechanism and template matching strategy to process semantic information and filter out dynamic feature points. Additionally, we replace the RANSAC algorithm with PROSAC to eliminate feature point mismatches and improve matching performance. Experimental evaluations on the public TUM RGB-D dataset demonstrate that compared with the original ORB-SLAM3 and the more advanced DynaSLAM algorithm, our improved method achieves superior localization accuracy and robustness in dynamic environments. The system maintains a frame rate exceeding 45 fps without GPU acceleration, ensuring real-time performance.