【摘要】 【目的】基于多视角图像的建筑物三维重建技术在城市规划和灾害模拟等领域具有广泛的应用价值。然而，现有方法在视觉效果优化的同时，往往难以兼顾几何精度、细节还原能力以及对复杂场景的适应性。针对复杂场景下建筑物高精度三维重建的挑战，本文提出了一种基于3D高斯溅射（3D Gaussian Splatting,3DGS）引导的多视图建筑物三维重建方法，旨在较好地捕捉建筑物的几何结构与细节特征，生成兼具高几何精度和高细节还原度的三维模型。【方法】该方法的核心流程分为3个阶段：首先，通过运动恢复结构（Structure from Motion,SfM）从多视角图像数据集中生成建筑物场景稀疏点云，再利用3D高斯溅射技术进行点云平滑和补全，生成连续的3D高斯建筑点云；其次，针对复杂建筑表面设计了3D高斯建筑点云优化策略，通过多尺度高斯表示，并基于局部曲率优化的正则化项，使3D高斯建筑点云更精确地贴合建筑物表面，提升模型的平滑度和视觉一致性；最后，采用Poisson重建从优化后的3D高斯建筑点云生成建筑物初始网格，并结合有符号距离场（Signed Distance Field,SDF）细化表面结构，进一步提高几何精度和细节保真度。基于本文提出的方法，在Tanks and Temples数据集中的Barn场景、ArcGIS公司提供的Small Buildings数据集和自采集的Tower数据集上开展实验，并与Colmap、Neuralangelo和SuGaR等方法进行比较。【结果】在几何精度方面，本文方法在Barn数据集上的F1分数（F1 Score）达到0.46，点到网格距离（Point-to-Mesh Distance）为0.049，均优于对比方法；在渲染质量方面，峰值信噪比（Peak Signal-to-Noise Ratio,PSNR）和结构相似性（Structural Similarity Index,SSIM）指标在Barn数据集上分别为27.83和0.92，在Small Buildings数据集上分别为29.67和0.94，在Tower数据集上分别达到32.69和0.96，均优于对比方法。【结论】本文方法在复杂场景下的建筑三维重建的几何准确性、细节保留能力和渲染质量方面较对比方法有较好提升，能够有效地恢复建筑物整体几何结构，并实现稳定的渲染性能。更多还原

【Abstract】 [Objectives] 3D reconstruction of buildings using multi-view images has important practical applications in a variety of fields such as surveying, visualization and urban management. Traditional image-based 3D reconstruction methods capture the 3D structure of a building by fusing image data from different perspectives to generate a 3D model of the building. However, it is often difficult for existing methods to take into account the geometric accuracy, detail restoration ability, and adaptability to complex scenes while optimizing the visual effect. Aiming at the challenges of high-precision 3D reconstruction of buildings in complex scenes, this paper proposes a 3D Gaussian Splatting-guided multi-view 3D reconstruction method for buildings, which can effectively capture the geometric structure and detailed features of buildings and generate 3D models with high geometric accuracy and superior detail preservation. [Methods] The core workflow of the proposed method consists of three stages: Firstly, Structure from Motion(SfM) is utilized to generate a sparse point cloud of the building scene from the multi-view image, and 3D Gaussian Splatting is utilized for point cloud smoothing and completion to generate a continuous 3D Gaussian point cloud. Secondly, a 3D Gaussian point cloud optimization strategy is designed for complex building surfaces, through multi-scale Gaussian representation and regularization terms based on local curvature optimization, so that the 3D Gaussian point cloud fits the building surface more accurately and enhances the smoothness and visual consistency of the reconstructed model. Finally, Poisson reconstruction is used to generate the initial mesh of the building from the 3D Gaussian point cloud, and combined with Signed Distance Field(SDF) to refine the surface structure to further improve the geometric accuracy and detail fidelity. Based on the proposed method, the experiments are validated on the Barn scene in the Tanks and Temples dataset, the Small Buildings dataset provided by ArcGIS and the self-collected Tower dataset, and the experimental results are compared with the Colmap, Neuralangelo, and SuGaR methods. [Results] In terms of geometric accuracy, the proposed method has an F1 Score of 0.46 and a Point-to-Mesh Distance of 0.049 on the Barn dataset, which is better than the comparison methods. In terms of render quality, the Peak Signal-to-Noise Ratio(PSNR) and Structural Similarity Index(SSIM) reach 27.83 and 0.92 on the Barn dataset, 29.67 and 0.94 on the Small Buildings dataset, and 32.69 and 0.96 on the Tower dataset, which are both better than compared methods. [Conclusions] The proposed method demonstrates significant improvements over comparison methods in terms of geometric accuracy, detail preservation and rendering quality in complex scenes, and is able to effectively recover the overall geometric structure of buildings and achieve stable rendering performance.更多还原