Visual Localization Using 3D Gaussian Splatting Representation for Mobile Robots With Geometric Feature Correspondences Synthesis

Abstract

Achieving visual localization with excellent interactive performance is challenging for mobile robots. Based on real-time photo-realistic view synthesis, 3D Gaussian splatting (3DGS) representation has demonstrated vast potential for robots engaging with the physical world. In this work, we propose a novel coarse-to-fine visual localization method named L3DGS based on the 3DGS radiance field representation. Particularly, during the coarse stage, we exploit novel views synthesized by the pretrained 3DGS map to create geometric feature correspondences to perform geometric alignment. Then, we integrate both geometric and photometric alignment to refine the camera pose. Unlike previous radiance field-based approaches, we leverage geometric feature correspondences and the innovative 3DGS map to improve the localization accuracy. In our experiments, we evaluate the proposed method across two real-world indoor and outdoor datasets. Consequently, compared to the baselines, the proposed method achieves competitive or superior experimental results.