NSDM-SLAM: Non-blocking Semantic Detection and Mask Propagation for Robust Visual SLAM in Dynamic Environments

To resolve the conflict between the high computational overhead of high-precision semantic segmentation and the real-time demands in existing dynamic SLAM methods, this paper proposes a non-blocking dynamic SLAM algorithm, NSDM-SLAM. Built upon ORB-SLAM3, our approach introduces an independent semantic segmentation thread and enhances the original tracking thread. The semantic thread employs a YOLO model to detect potential dynamic objects and generate masks in real-time, while the tracking thread utilizes an innovative mask propagation algorithm to obtain precise dynamic regions in the current frame. Unlike existing methods, our algorithm distinguishes between dynamic and static components of the same object (e.g., the upper and lower body of a seated person engaged in conversation), thereby preserving more reliable static feature points for localization. The system incorporates depth, geometric constraints, and a probabilistic fusion mechanism to construct more accurate dynamic masks. This non-blocking design ensures that the tracking thread does not wait for semantic processing results, thereby improving the system's real-time performance. Experiments on the TUM RGB-D dataset demonstrate that, compared to the baseline ORB-SLAM3, our method reduces the Absolute Trajectory Error by up to 97.7% in highly dynamic sequences and achieves accuracy comparable to other state-of-the-art methods. Benefiting from the non-blocking design, the system achieves an average processing speed of 34.5 FPS without GPU acceleration. In real-world tests with a static camera, trajectory drift under dynamic interference was reduced by over 98% compared to ORB-SLAM3.