DRG-SLAM: Dynamic-Aware Gaussian Splatting SLAM for RGB-D Scenes
Abstract
Dynamic objects remain a major challenge for dense visual simultaneous localization and mapping (SLAM), as they can corrupt both camera tracking and map reconstruction. Although 3D Gaussian Splatting (3DGS) provides an explicit and efficient scene representation for high-quality rendering, most 3DGS-based SLAM methods still rely on the static-scene assumption and are therefore sensitive to moving objects. This letter presents a dynamic-aware 3DGS-SLAM framework for robust pose estimation and static scene reconstruction in RGB-D dynamic environments. Instead of directly assigning dynamic labels according to semantic categories, our method uses instance masks only as object-level candidates. Dynamic objects are detected by combining depth inconsistency with ego-motion-compensated displacement variation, and a sliding-window strategy is used to accumulate temporal evidence for stable motion judgment. Furthermore, a geometry-guided boundary refinement module is introduced to suppress unreliable observations around dynamic object boundaries. Based on the refined static regions, camera poses and the Gaussian map are optimized using only reliable static observations. Experiments on the TUM RGB-D and BONN RGB-D datasets demonstrate that the proposed method achieves better tracking accuracy and reconstruction quality than representative Neural Radiance Field (NeRF) and 3DGS-based SLAM methods in challenging dynamic scenes.