基于三维高斯泼溅的港口场景高效轻量化三维重建

High-Efficiency Lightweight 3D Reconstruction of Port Scenes Based on 3D Gaussian Splatting

摘要: 针对智慧港口数字孪生建设对高精度、轻量化三维重建的需求，现有激光雷达、倾斜摄影和神经辐射场等传统方法在港口场景中存在的成本高、建模慢、动态适配差等瓶颈，提出一种基于三维高斯泼溅的港口场景三维重建方法。构建涵盖稀疏视角鲁棒建模与轻量化部署的三维高斯泼溅重建框架，采用改进运动恢复结构初始化、港口结构约束的损失函数和分级自适应密度控制等关键技术。在港口场景的实验表明，该方法单场景建模耗时7.1 min，较基线仅增加4.4%，较基础3DGS在重建精度、渲染效率及轻量化部署方面优势显著，可有效支撑智慧港口数字孪生建设对高精度、轻量化三维重建的需求。

Abstract: Aiming at the requirements of high-precision, high-efficiency and lightweight 3D reconstruction for smart port digital twin construction, this paper analyzes the bottlenecks of traditional methods such as LiDAR, oblique photography and NeRF in port scenes, including high cost, slow modeling and poor dynamic adaptability. A port scene 3D reconstruction method based on 3D Gaussian Splatting is proposed. A 3D Gaussian Splatting reconstruction framework covering sparse-view robust modeling and lightweight deployment is constructed. Key technologies including improved Structure-from-Motion initialization, multi-constraint loss function with port structural constraints, and hierarchical adaptive density control are adopted. Experiments on port scenes show that the single-scene modeling time is only 7.1 min, only 4.4% more than the baseline. Compared with baseline 3DGS, the proposed method has significant advantages in reconstruction accuracy, rendering efficiency and lightweight deployment, which can effectively support high-precision and lightweight 3D reconstruction for smart port digital twin construction.