Real-time RGB-D SLAM with Gaussian scene representation for dynamic environments

Abstract This paper presents a lightweight semantic front-end for improving registration-driven Gaussian RGB-D SLAM in dynamic scenes. In dynamic environments, dynamic objects significantly degrade the robustness of Gaussian-based RGB-D SLAM systems by introducing inconsistent geometric observations during registration and map fusion. To address this problem, we propose an input-level dynamic removal framework for Gaussian-based SLAM that integrates real-time instance segmentation with a registration-driven Gaussian mapping pipeline. To maintain real-time performance under limited computational resources, the segmentation module is executed periodically and the most recent mask is reused for intermediate frames. Experimental results on public RGB-D benchmarks and real-world indoor sequences show that the proposed method improves trajectory robustness in dynamic environments and reduces dynamic-region contamination compared with the baseline neural explicit SLAM pipeline.