Abstract: [Purposes] Aiming at the problems of complex street layout,intricate architectural structures and variable illumination in waterfront areas of coastal historic villages,which easily cause holes,rough edges and noise artifacts in large-scale 3D reconstruction,this paper takes Houguan village in Fuzhou as the research area and proposes an improved large-scale realistic 3D reconstruction framework based on CityGaussianV2. [Methods] Multi-view images are acquired via five-direction oblique photography using UAV.The prior weight of building edges is adopted to guide the splitting candidate optimization,so as to improve the reconstruction accuracy of building contours and component boundaries.Adaptive kernel density estimation（AKDE） is introduced for point cloud denoising to suppress outliers and noise artifacts. [Findings] Experimental results indicate that the overall RMSE values of Octree-GS,Mip-Splatting,LightGaussian and CityGaussianV2 are 4.2,4.0,3.7 and 4.6 cm respectively.The overall RMSE of the proposed method is 2.8 cm,with an error reduction ranging from 24.3% to 39.1% compared with all comparison methods. [Conclusions] The proposed method can provide a technical solution for large-scale and high-precision 3D modeling of historic villages,and satisfy the demands of digital archiving and smart tourism．

Key words: real-scene 3D modeling, unmanned aerial vehicle oblique photography, building edge prior, 3D point cloud denoising, architectural conservation