综合数据驱动与模型驱动的机载LiDAR数据复杂建筑建模方法

The three-dimensional building model is one of the core fundamental information under the rapid growth of city area, which plays an important role in city planning, city management, digital city, etc. In this research, an automatic hybrid-driven method for reconstructing 3D geometrical building models is developed through the combination of the data-driven and model-driven methods based on the discrete-return and full-waveform airborne light detection and ranging (LiDAR) data. The complicated building rooftop is first reasonably segmented into non-overlapping quadrilaterals with common edges under semantic constrains. Then the quadrilaterals are reconstructed by matching the primitive library objects through the optimization theory and statistical analysis methods. Finally, an integrated building model can be obtained by composing those quadrilateral sub-models together. In the hybrid-driven framework, an improved random sample consensus (RANSAC) algorithm based on data-driven is used to correctly segment the rooftop patches of each building or group of connected buildings which often lie in shantytowns or along the urban street. The algorithm successfully maintains the topological consistency among primitives and can avoid under-segmentation and over-segmentation. The robust model-driven matching method abandons the unreasonable assumption that all the building rooftops are polyhedral in the former researches. By fusing the data-driven and model-driven framework, we can reconstruct not only the polyhedral building models but also the complicated building models which are composed of the plane, cylinder, cone, sphere, irregular curved surface or any combinations. The results of this research would expand and improve the existing building modeling framework and have a higher academic value and application potential.

作为城市核心三维基础地理信息的三维建筑模型在城市规划、城市管理和数字城市等行业发挥了越来越重要的作用。本研究拟融合数据和模型驱动的思想，基于LiDAR离散点云和全波形LiDAR数据，研究复杂建筑屋顶自动重建的技术与方法。首先利用数据驱动，在特定语义约束下分割复杂建筑屋顶为分割单元，然后对分割单元采用模型驱动，并借助最优化理论和统计分析方法匹配参数模型基元库，最终组合所有分割单元模型，完成复杂建筑屋顶的建模。改进的基于数据驱动屋顶面片分割RANSAC算法，保证了分割面片拓扑关系，克服了过分割或欠分割的缺陷，能够合理分割复杂建筑或连体建筑屋顶；基于模型驱动的匹配方法摒弃了传建模方法中认为"屋顶是由平面构成的多面体模型"的不合理假设，可构建平面、圆柱、圆锥和圆球等规则几何元素或不规则曲面，或者其组合构成的复杂屋顶。本课题将拓展LiDAR数据建筑方法，实现大面积三维建筑模型的快速更新及工程化应用。

建筑是组成城市最重要的核心单元，也是建立城市三维地理信息系统的基础，因此快速准确地获取城市三维建筑数据并建立三维建筑模型（Digital Building Model, DBM）对于城市规划、城市管理、智能交通、应急救助、房地产展示、旅游推介、数字城市、灾害模拟分析和基于位置的服务（Location Based Service, LBS）等诸多领域具有较高的应用价值。基于此背景，本项目研究了如何从波形数据中结合数据驱动和模型驱动构建三维建筑模型的方法。具体而言，本项目研究了设计正则约束和数据约束下的多尺度全波形分解优化算法。通过引入多类核密度函数，构建核密度函数库，拟合后向散射波形中形态各异的回波。同时统计多平台波形数据的分布特征，研究先验知识约束下的波形数据拟合方法，解决单纯依靠核密度函数造成的波形过拟合问题。然后综合分析核密度函数和先验知识，探索结合扫描场景和多尺度波形数据的核密度函数和先验知识的动态自由组合策略，构建数据约束和先验知识约束下的优化模型，在提高波形分解和拟合精度的同时，全面提取回波的几何和辐射特征，为后精确的建模奠定了基础。针对建筑几何建模，提出了基于数据驱动的机载建筑屋顶建模方法及融合模型驱动的混合建模方法。通过在二维空间维持屋顶面片间的拓扑关系，将三维建模问题转化为维护二维屋顶平面的拓扑问题，继而进一步兼容屋顶面片在高程维度的拓扑一致性。然后，提出了建筑分割单元的匹配、规则曲面几何元素的匹配、和复杂分割单元屋顶的简化建模三种策略实现了数据驱动和模型驱动混合建模。本项目的实施对于综合利用多类型激光雷达数据，提升智慧城市的数字化管理能力，完善激光雷达理论与方法具有重要的理论和实践意义。