Abstract:

To address the challenges of low computational efficiency, slow convergence, and limited accuracy in large-scale 3D scene registration, a 3D Gaussian splatting (3DGS) registration method integrating GPS prior information was proposed. Spatial position priors provided by GPS were utilized to establish initial alignment through coordinate system transformations, narrowing the registration search space. Dense point cloud models were efficiently reconstructed by combining 3DGS technology. High-precision alignment was achieved through a two-stage optimization of GPS coarse registration and fine registration. Experiments demonstrate that the GPS-assisted method reduces translation errors by 25%~50% and increases success rates to 98% in vegetation-covered and open-square scenarios. This method provides efficient technical support for large-scale scene reconstruction in smart cities and digital twin applications.

Key words: GPS prior information, 3D Gaussian splatting(3DGS), point cloud registration, large-scale scene reconstruction, coordinate system transformation