A Dynamic Dense SLAM Algorithm Based on 3D Gaussian Splatting

Currently, dense SLAM algorithms based on 3DGS achieve impressive mapping results. However, most of them neglect the removal of dynamic points, leading to suboptimal performance in dynamic environments. To address this issue, a 3DGS-based approach enhanced with mask-based dynamic point filtering is proposed to improve mapping accuracy. Another scenario is that dynamic elements in a scene include not only moving individuals but also stationary objects like chairs and everyday items that can be displaced for which the current algorithms cannot obtain satisfactory results. Consequently, a method for detecting various object movements is proposed, combined with a dynamic object filtering strategy that integrates reprojection error with the intersection-over-union (IoU) ratio. To further enhance efficiency, the tracking process is replaced by an adaptive keypoint-based method, significantly reducing computational costs. Experimental results demonstrate the robustness and real-time performance of the proposed Gaussian Dynamic SLAM(GSD-SLAM) system in dynamic environments.