RGS-SLAM: Robust Gaussian Splatting SLAM with Image Reconstruction in Dynamic Scenes
Simultaneous Localization and Mapping (SLAM) enables autonomous navigation and mapping for mobile robots, but most visual SLAM systems assume static environments and fail in the presence of dynamic objects, leading to localization drift and distorted maps. To address this, we propose RGS-SLAM, a robust Gaussian Splatting–based SLAM system that integrates YOLOv8 for dynamic object detection and LaMa inpainting to restore occluded regions, ensuring stable tracking and dense map construction. The system employs 3D Gaussian Splatting for efficient scene representation and photorealistic rendering. Experiments on the TUM RGB-D dynamic dataset demonstrate that RGS-SLAM reduces trajectory error by up to 91.6% compared with ORB-SLAM3 and generates high-quality dense maps under dynamic conditions. Moreover, the system runs efficiently on the NVIDIA Jetson AGX Orin, highlighting its feasibility for real-world edge applications.