Semantic perception 3DGS-SLAM system for dynamic scenes

Abstract The integration of 3D Gaussian Splatting (3DGS) with visual simultaneous localization and mapping (SLAM) enables the construction of high-fidelity scene maps. However, most existing methods are designed for static environments, exhibiting significant performance degradation in dynamic scenes. To address this critical limitation, this paper proposes a novel semantic-aware SLAM system (GSV-SLAM) for dynamic scenes. Specifically, GSV-SLAM first deploys a YOLOv11-based semantic segmentation network to identify and remove predefined dynamic objects from input frames. Subsequently, it extracts ORB features exclusively from static background regions. This optimized processing sequence effectively reduces redundant feature computations while maintaining real-time performance. Second, an adaptive geometric consistency constraint strategy is employed to further suppress dynamic feature points on potential moving objects and reconstruct occluded background information caused by moving objects. Finally, a Gaussian map is constructed using optimized camera poses and static point clouds to achieve high-quality 3D reconstruction and new viewpoint rendering. Experimental results on the TUM and BONN datasets demonstrate that the proposed method achieves significantly lower absolute and relative trajectory errors compared to ORB-SLAM3. Real-world experiments further validate the system’s robustness in scenarios with strong dynamic disturbances.