SDD-SLAM: Semantic-Driven Dynamic SLAM With Gaussian Splatting

Abstract

Recently, significant advancements have been made in 3D Gaussian Splatting SLAM for dynamic environments. However, most existing methods primarily address active dynamic objects, such as people and vehicles, and fail to account for the impact of passive dynamic objects on localization and mapping. This results in the presence of numerous artifacts left by dynamic objects in the scene, which diminishes the accuracy of pose estimation. To address these challenges, we propose SDD-SLAM, a semantic-driven SLAM system based on 3D Gaussian Splatting. Extensive experiments conducted on the TUM and BONN datasets demonstrate that the proposed methods, including refined mask expansion, edge noise filtering, object-level dynamic object removal based on semantic Gaussians, and object-level density control strategy for Gaussian ellipsoids, significantly enhance the accuracy of camera pose estimation and the quality of map reconstruction in dynamic environments, achieving state-of-the-art performance.