VLGS-SLAM: Visual–Lidar Three-Dimensional Gaussian Splatting Simultaneous Localization and Mapping
Recent studies highlight the effectiveness of 3D Gaussian splatting (3DGS) in visual simultaneous localization and mapping (SLAM) systems, which work well indoors but struggle in large-scale outdoor environments. Typically, lidar data are used to address this issue; however, current multi-modal SLAM systems use 3DGS mainly for mapping, leaving its potential to enhance tracking unexplored. In this paper, we present VLGS-SLAM, a novel visual–lidar SLAM pipeline that leverages lidar data and 3DGS for both pose estimation and mapping. Our approach integrates lidar points as 3D Gaussian primitives, ensuring precise scene geometry and reducing pose estimation errors caused by floating Gaussians. To enhance tracking performance, we apply regularization to Gaussian scaling, which constrains the shape of each Gaussian ellipsoid. For loop closure, we combine image similarity with lidar cloud distance to effectively detect and close loops. Our experiments demonstrate that VLGS-SLAM achieves state-of-the-art accuracy in the 3DGS-based SLAM field, outperforming many traditional SLAM algorithms