ReLoc-3DGS: A LiDAR and Vision-Based Relocation Framework Using 3D Gaussian Splatting for Seamless Indoor-Outdoor Environments

This paper presents a novel 3D map representation and visual relocalization framework based on 3D Gaussian Splatting. By integrating LiDAR, camera, and IMU data, our method constructs a visually realistic environmental representation with strong generalizability, making it applicable to a wide range of Autonomous Mobile Robots (AMRs).We initialize the training of the 3D Gaussian Splatting map using sparse point clouds generated from odometry, resulting in a semantically rich map. To improve localization performance, we propose a two-stage coarse-to-fine relocalization pipeline: first, we distill the 3D Gaussian Splatting map into a custom-designed neural network, XLOC, to achieve end-to-end image-to-position coarse localization; then, we apply a localized particle filter to refine the pose estimation. Experimental results demonstrate that our framework achieves efficient and robust relocalization across various unstructured indoor and outdoor environments, showing strong adaptability to diverse AMR platforms.