TLS-SLAM: Gaussian Splatting SLAM Tailored for Large-Scale Scenes

Abstract

3D Gaussian splatting (3DGS) has shown promise for fast and high-quality mapping in simultaneous localization and mapping (SLAM), but faces convergence challenges in large-scale scenes across three key aspects. Firstly, the excessive Gaussian points in 3DGS models for large-scale scenes make the search space of the model optimization process more complex, leading to local optima. Secondly, trajectory drift caused by long-term localization in large-scale scenes displaces Gaussian point cloud positions. Thirdly, dynamic objects commonly found in large-scale scenes produce numerous noise Gaussian points that disrupt gradient backpropagation. We propose TLS-SLAM to address these convergence challenges. To ensure large-scale scene map optimization attains the global optimal, we use scene memory features to encode and adaptively build sub-maps, dividing the optimization space into subspaces, which reduces the optimization complexity. To reduce trajectory drift, we use a pose update method guided by semantic information, ensuring accurate Gaussian point cloud creation. To mitigate the impact of dynamic objects, we utilize 3D Gaussian distributions to accurately extract, encode, and model dynamic objects from the scene, thereby avoiding the generation of noise points. Experiments on four datasets show that our method achieves strong performance in tracking, mapping, and rendering accuracy.