Research on a motion-awareness guided semantic dynamic SLAM algorithm

Abstract Visual simultaneous localization and mapping (SLAM) systems in dynamic environments suffer from degraded localization accuracy due to moving objects, creating a bottleneck for practical deployment. To address this issue, this paper proposes a motion-awareness guided semantic dynamic SLAM algorithm. The core of the algorithm lies in establishing a coarse-to-fine dynamic feature processing pipeline: Firstly, optical flow is utilized for efficient motion awareness, achieving preliminary localization and rapid screening of dynamic regions within the image. This motion information serves as a decision-making basis to intelligently control the activation of the subsequent semantic analysis module. Subsequently, for the identified motion regions, a lightweight improved YOLOv11-GC2f model is activated for precise instance segmentation. By incorporating GhostConv and C2fPSA modules, this model significantly enhances inference efficiency while maintaining segmentation accuracy. Finally, the algorithm integrates the results of motion awareness and semantic segmentation, employing a union operation to maximally eliminate dynamic feature points, thereby ensuring that camera pose estimation relies solely on reliable static scene structures. Experimental validation demonstrates marked improvements in both accuracy and efficiency. On the TUM RGB-D dataset, the proposed algorithm reduces the absolute trajectory error by 81.17% in highly dynamic sequences compared to ORB-SLAM3. Additional evaluation on the Bonn dynamic object dataset shows consistent robustness across diverse controlled scenarios. Crucially, the motion-aware triggering mechanism substantially reduces computational overhead compared to per-frame semantic processing, enabling real-time operation while maintaining these accuracy gains. By effectively guiding semantic computation through motion awareness, this research provides an efficient and robust solution for achieving precise, long-term operation of robots in dynamic environments.