Core Viewpoint - The article presents a unified LiDAR-visual system that addresses geometric inconsistencies in Gaussian splatting for robotic applications, successfully combining Gaussian splatting with Neural Signed Distance Fields (NSDF) to achieve geometrically consistent rendering and reconstruction [52]. Group 1: Unified LiDAR-Visual System - The proposed system aims to utilize registered images and low-cost LiDAR data to reconstruct both the appearance and surface structure of scenes under arbitrary trajectories [5][6]. - The importance of Gaussian initialization in achieving good structure is emphasized, highlighting its role in the optimization process [22]. Group 2: Geometric Regularization - The article discusses the introduction of geometric regularization into the 3D Gaussian Splatting (3DGS) framework to address geometric inconsistencies that manifest as rendering distortions [3][6]. - It suggests that depth cameras and LiDAR can provide direct structural priors, which can be integrated into the 3DGS framework for improved geometric regularization [3]. Group 3: Methodology - The overall process includes three stages: training a Neural Signed Distance Field (NSDF) using point clouds, initializing Gaussian primitives from the NSDF, and optimizing both Gaussian primitives and NSDF through SDF-assisted shape regularization [8][6]. - The use of 2D Gaussian splatting to represent 3D scenes is detailed, with each disk defined by parameters such as center point, orthogonal tangent vectors, scaling factor, opacity, and view-dependent color [10]. Group 4: Experimental Results - The proposed method demonstrates superior reconstruction accuracy and rendering quality across various trajectories, as evidenced by extensive experiments [52]. - Quantitative results indicate that the method outperforms existing techniques in metrics such as C-L1, F-Score, SSIM, and PSNR across multiple datasets [46][49]. Group 5: Limitations and Future Work - The method exhibits limitations in extrapolating new view synthesis capabilities, suggesting a need for further exploration of advanced neural rendering techniques to address this limitation [53].