LiDAR-Guided 3D Gaussian Splatting with Differentiable UDF-Based Regularization for Mine Tunnel Reconstruction

Underground mine tunnels are often characterized by extremely uneven illumination, weak surface textures, and frequent dynamic interference, which severely undermine multi-view photometric consistency and easily induce floating artifacts and spatial divergence in conventional vision-based 3D Gaussian Splatting (3DGS). To address these issues, we propose a LiDAR-guided 3DGS framework for underground tunnel reconstruction based on dynamic-object removal and differentiable unsigned distance field (UDF) regularization. First, a dynamic foreground removal strategy with background restoration is introduced to remove transient foreground disturbances and restore static supervision consistency. Second, LiDAR point clouds are leveraged to initialize Gaussian primitives with a reliable geometric skeleton in weak-texture regions. More importantly, LiDAR priors are further converted into a differentiable UDF field and serve as a persistent geometric constraint. A dual-track mechanism is designed, where continuous geometric attraction pulls mildly deviated Gaussians back toward the physical surface and periodic out-of-bound culling removes severely drifting primitives. Experiments on real underground tunnel and chamber scenes show a clear scene-dependent behavior of the proposed method. In the tunnel scene, the method achieves the best SSIM together with competitive PSNR and LPIPS, while also reducing redundant out-of-bound primitives and improving geometric cleanliness. In the chamber scene, however, its advantages under global full-reference metrics are less evident. These results suggest that the proposed LiDAR-guided and differentiable UDF-regularized framework is particularly beneficial for weak-texture tunnel environments, while further improvement is still needed for chamber scenes with more complex appearance variations.