测绘通报 ›› 2026, Vol. 0 ›› Issue (10): 80-85,91.doi: 10.13474/j.cnki.11-2246.2026.1012

• 学术研究 • 上一篇    

一种空地影像协同的三维高斯泼溅建模方法

崔扬1,2, 陶迎春1,2, 刘光1,2, 刘鹏1,2, 安智明1,2, 王琳1,2   

  1. 1. 北京市测绘设计研究院, 北京 100038;
    2. 城市空间智能感知与数字治理北京市重点实验室, 北京 100038
  • 收稿日期:2026-01-04 发布日期:2026-11-06
  • 通讯作者: 陶迎春。E-mail:527359590@qq.com
  • 作者简介:崔扬(1990—),女,博士,高级工程师,主要研究方向为三维智能处理。E-mail:cuiyangmailbox@163.com
  • 基金资助:
    国家自然科学基金(42201483);北京市教育委员会科学研究计划(KM202410016007)

A 3D Gaussian splatting modeling method based on aerial and ground image collaboration

Cui Yang1,2, Tao Yingchun1,2, Liu Guang1,2, Liu Peng1,2, An Zhiming1,2, Wang Lin1,2   

  1. 1. Beijing Institute of Surveying and Mapping, Beijing 100038, China;
    2. Beijing Key Laboratory of Urban Spatial Intelligent Sensing and Dagital Geovernance, Beijing 100038, China
  • Received:2026-01-04 Published:2026-11-06

摘要: [目的] 航拍与近地面影像的无缝化三维建模是实景建模中的重大挑战,本文以北京法海寺与白塔寺为研究对象,提出了一种基于空地影像协同的三维高斯泼溅(3DGS)建模方法,旨在解决空地视角剧烈变化场景的高质量重建。[方法] 首先,采用近地面多模态异构设备阵列采集策略,实现复杂场景全方位、高精度近地面影像获取;然后,提出分步解算刚性约束方法,完成空地一体化联合空中三角测量;最后,采用分块并行训练、多源相机视角平衡、生成式视角增强与时空一致性清洗的组合建模策略,重建高保真、无缝化三维高斯泼溅模型。[结果] 结果表明,该方法在模型完整性、几何精度反视觉细节方面均优于主流建模方法,峰值信噪比(PSNR)提升1.143 7 dB,结构相似性(SSIM)提升0.022 3,学习感知图像块相似度(LPIPS)降低0.018 7。[结论] 本文方法有效解决了大视角变化下的高质量三维建模难题,为文化遗产数字化保护与城市精细化管理提供了新的技术途径。

关键词: 空地协同, 多模态数据, 空中三角测量, 三维高斯泼溅, 实景三维建模

Abstract: [Purposes] 3D modeling of aerial and ground image collaboration is a significant challenge in real-scene modeling.This paper focuses on Fahai temple and White Pagoda temple in Beijing as research objects and proposes a 3D Gaussian splatting (3DGS)method based on aerial and ground image collaboration,which aiming to solve the high-quality reconstruction of scene with drastic changes in aerial and ground perspectives. [Methods] Firstly,the ground multi-modal heterogeneous equipment array acquisition strategy is implemented to achieve comprehensive and high-precision near-ground images of complex scenes.Secondly,a stepwise rigid constraint resolution method is employed to accomplish integrated aerial-ground joint aerial triangulation.Finally,a combined modeling strategy integrating block-based parallel training,multi-source camera viewpoint balancing,generative viewpoint augmentation,and spatio-temporal consistency filtering is utilized to reconstruct a high-fidelity and seamless 3D Gaussian splatting model. [Findings] The results demonstrate that the proposed method outperforms mainstream modeling approaches in terms of model completeness,geometric accuracy and visual details.Specifically,the peak signal-to-noise ratio (PSNR)is increased by 1.143 7 dB,the structural similarity index measure (SSIM)is improved by 0.022 3,and the learned perceptual image patch similarity (LPIPS)is reduced by 0.018 7. [Conclusions] The proposed method effectively addresses the challenge of high-quality 3D modeling under large viewpoint variations,which offers a new technological pathway for digital preservation of cultural heritage and fine-grained urban management.

Key words: aerial and ground collaboration, multi-modal data, aerial triangulation, 3D Gaussian splatting, real-scene 3D model

中图分类号: