面向建筑物精细建模的倾斜摄影测量理论与方法

Automatic building reconstruction using oblique photogrammetry technique has encountered severe problems from limited types of primitives in the model-driven approach and suffered from evident noises in dense matched point clouds, which lead to conspicuous falloff in the detail and precision of the reconstructed model. Aiming at improving the details of reconstructed buildings, this project incorporates geometrical and texture knowledge with the matching of oblique images, from which improved dense matched point clouds are expected. Furthermore, four classes of structure knowledge are refactored from the limited primitive library, including coplanarity, parallelism, orthogonality and symmetry, which are more general compared to the fixed primitive geometry and injected into the procedures of plane fitting and building reconstruction to alleviate the problem caused by noises of point clouds. The key objectives of this project include: 1)Bounds constrained optimization for feature point matching of oblique images, which convert the initial geometrical information into bounding constraints and seamlessly integrated into the matching procedure to improve reliability and accuracy of aerial triangulation; 2) Texture-aware dense image matching of oblique images, which are used to reduce the noises of photogrammetric point clouds at low texture area as well as to preserve features; 3) Point clouds classification using multilevel semantic relationships, which grants semantic information to the original point clouds and enables the extraction of individual building from the point clouds; 4) Plane extraction and topological reconstruction of the building using structure knowledge, which creates detailed polygonal building models from point clouds and 5) prototype software development and evaluations using typical oblique images from practical productions and benchmark datasets. This project is devoted to dramatically improve the automatic processing of oblique images to create detailed building models.

面向建筑物精细三维建模，突破现有倾斜摄影测量方法受限于固定基元模型类别，难以适用于错综复杂的建筑物形状结构，且受密集匹配点云噪声和精度影响，三维建模自动化程度低且模型质量较差等理论瓶颈。主要研究内容包括：1）边界约束的倾斜影像特征匹配，将传统最小二乘匹配转换为一个边界约束优化问题，显著提升收敛率；2）纹理感知的倾斜影像密集匹配，保特征并抑制低纹理区域噪声；3）多级语义关联的建筑物识别与提取，实现建筑物可靠的单体化分割；4）结构知识引导的可靠面片提取及拓扑重建，自动恢复面片交线和交点等精细的拓扑结构；5）研发自主知识产权的软件原型系统，并选择典型的城市地区倾斜影像进行实验验证。本项目的研究将为大规模精细化的三维城市自动建模提供有效的理论依据和关键技术支撑，促进倾斜摄影测量理论方法的发展。

精细的建筑物实景三维模型是智慧城市建设与城市安全保障和VR/AR等智能应用的关键基础信息，倾斜摄影测量是实景三维城市建模的主要手段之一，但已有倾斜摄影测量三维建模方法由于自动化程度低、模型质量较差、且严重依赖人工交互操作，难以满足大中小各种规模城市全域范围高效的建筑物精细三维建模需求。为此，项目组较系统地突破了面向建筑物精细三维建模的倾斜摄影测量共性基础难题，包括影像几何变形与辐射质量差异大导致的特征可靠匹配与密集匹配点云噪声明显、点云语义标识缺失与建筑物模型结构错综复杂导致的单体化建筑物精细三维模型重建等，初步形成了面向建筑物精细建模的倾斜摄影测量成套理论方法体系，为实景三维中国和智慧城市等国家重大工程提供了关键技术支撑；主要创新点有：（1）宽基线倾斜影像可靠特征匹配方法，有效克服倾斜影像几何变形和辐射质量问题，切实提高倾斜影像匹配效率；（2）顾及特征结构的倾斜影像密集匹配方法，结合影像纹理信息自适应调整密集匹配参数，显著降低点云噪声；（3）复杂场景语义智能感知与建筑物提取方法，将点云数据的多层次语义关联显式的引入点云分类，获取可靠的密集匹配点云语义信息，实现建筑物单体化提取；（4）多细节层次建筑物精细模型高效重建方法，数据驱动和模型驱动加结构知识引导的建筑物精细建模方法，实现建筑物精细三维模型自动重建。自主研发了一套倾斜摄影测量精细三维建模软件OSketch，商业化推广应用逾400家单位，在华为和阿里云举办的大赛中分别荣获冠军和亚军；相关成果荣获2021年度国家科技进步二等奖1项；共发表高水平论文37篇（其中，本领域顶刊 ISPRS 会刊论文10篇，荣获国际期刊学术论文奖励 1 项）；发明专利6项（授权4项），获得软件著作权2项；培养博士后3人、博士生14人、硕士生20人；项目负责人朱庆2017年被评为国家测绘地理信息局科技领军人才，项目骨干胡翰教授2021年被评为国家高层次青年人才，项目成员陈敏副教授2021年被评为高校GIS新锐。