Research on High-Precision 3D Modeling Technology for Campus Real-Scenes Based on NeRF

Aiming at the requirement of 3D modeling of medium-scale and high-complexity scenes in the campus, this study proposes a high-precision real scene modeling technology system based on neural radiation field (NeRF). Aiming at the deformation artifact problem of traditional oblique photography modeling in transparent material reconstruction and the limitation of insufficient stability of radiation field under dynamic illumination, a space-ground collaborative multi-source data fusion framework is constructed, which integrates five-lens oblique photography, ground-based laser scanning and backpack SLAM system, and achieves submillimeter spatial alignment through hash coding and iterative registration. A dynamic block modeling strategy is proposed, which trains independent NeRF models according to functional partition, combines progressive mesh optimization and U-Net feature fusion, and retains sub-centimeter microstructure features. A cascade super-resolution enhancement module is designed, and a texture refinement mechanism guided by light transmittance and density features is introduced to significantly improve the visual fidelity of high-frequency details. The experimental results show that the root mean square error (RMSE) of geometric reconstruction is $\text{1.2 cm}$, which is 65.7% lower than traditional method, the reflection error of glass curtain wall is reduced by 62%, the character recognition rate is improved to 98%, and the modeling efficiency is 85.7% shorter than that of oblique photography, which provides a high confidence three-dimensional base for precise space management and intelligent operation and maintenance of scientific campus.