PR-SLAM: Parallel Real-Time Dynamic SLAM Method Based on Semantic Segmentation

Abstract

SLAM (Simultaneous Localization and Mapping) is the core technology enabling autonomous exploration by mobile robots in unknown environments. While there have been numerous impressive SLAM systems developed, many of them are primarily based on the assumption of static environments, limiting their applicability in real-world settings. In order to enhance the robustness and accuracy of systems in dynamic real-world scenarios, we have introduced a parallelized real-time SLAM system called PR-SLAM, building upon the foundation of ORB-SLAM3. This algorithm introduces a dynamic probability update strategy within the semantic segmentation thread, completely decoupling the semantic segmentation thread from the tracking thread. Theoretically, the processing time per frame is solely dependent on the runtime of the tracking thread. Furthermore, we employ a geometric approach based on reprojection error to compensate for semantic gaps generated during semantic segmentation model inference. We have also designed a semantic optimization thread based on the dynamic probability of map points to optimize camera poses during semantic gaps. Finally, to reduce semantic gaps, we have performed lightweight modifications to SOLOV2. Comparative experiments were conducted against the state-of-the-art SLAM systems using the TUM dataset. The results indicate significant improvements in both accuracy and real-time performance for PR-SLAM. When compared to ORB-SLAM3, PR-SLAM achieved a remarkable 97.83% improvement in absolute trajectory accuracy and demonstrated an impressive 86.71% increase in runtime speed compared to DynaSLAM.