【作者】 王峰；

【导师】 朱庆；

【作者基本信息】 西南交通大学 ， 测绘科学与技术， 2020， 博士

【摘要】 多细节层次的建筑物三维模型包括了LOD-0到LOD-3四种典型的细节层次模型,面向不同层级的应用需求分别具有不同的几何和语义信息,特别是可以区分建筑物屋顶和立面结构的LOD-2模型构成了智慧城市的骨架结构,在城市规划建设与管理中的应用最为广泛。倾斜影像由于其立面可见性、表达直观性和多视冗余性,已经成为大规模城市LOD-2模型重建的主要数据源。针对已有倾斜摄影测量LOD-2重建方法存在的1)难区分和提不准的技术瓶颈难题;2)高噪声、遮挡严重以及语义信息缺失的数据质量问题引起的拓扑重构难题,导致LOD-2模型重建自动化程度低,模型质量差。为此,本文充分利用既有的众源建筑物轮廓数据,研究轮廓特征约束的倾斜摄影测量建筑物LOD-2模型重建方法,旨在有效破解多种类和形状各异的建筑物实体结构的识别、提取与多层次细节重建难题,具体研究内容如下:(1)既有建筑物轮廓数据与倾斜摄影测量点云的多基元配准方法既有建筑物轮廓数据与倾斜摄影测量点云数据来源以及表达形式的不同,不可避免的存在位置偏差,因此两者的精确配准极为重要。本文提出一种建筑物轮廓数据与倾斜摄影测量点云的面-线-点多基元配准方法,以平面特征为基元,从点云中提取建筑物竖直平面特征;以线特征为基元,构建线状结构特征,采用基于RANSAC的匹配方法,搜索建筑物轮廓数据与点云立面轮廓线特征的对应线状结构特征;以点为基元,对具有对应关系的线状结构特征的交点使用全局最小二乘优化策略,实现建筑物轮廓数据与点云的精确配准。(2)全局法线优化的倾斜摄影测量点云可靠平面提取方法城市建筑物模型通常较为复杂,但可以将其视为多种简单几何平面组合而成,因此从建筑物点云中提取可靠、精确的平面基元至关重要。然而,由于倾斜摄影测量点云高噪声的问题,传统区域生长平面提取方法难以提取完整的平面基元。为此,本文提出一种全局法线优化的倾斜摄影测量点云可靠平面提取方法。1)通过相似性聚类构建具有严格平面特征的超体素;2)引入最大平面支持区域概念,融合相邻体素构建最大平面支持区域;3)采用点云局部-全局一致性空间约束方法,将局部特征转化为全局特征,优化点云法线;4)最终使用最大平面支持区域引导的平面提取方法实现建筑物平面的准确提取。(3)边缘线立面结构引导的建筑物精细结构优化方法现有模型驱动方法的有限参数化基元模型,难以支持复杂环境建模,而数据驱动方法受制于点云噪声,即使最简单的平面基元也难以提得全、提得准,更无法准确恢复基元间的规则结构。为此,本文提出一种边缘线立面结构引导的建筑物精细结构优化方法。1)以众源的建筑物轮廓数据为约束,以准确提取的建筑物竖直平面为基础,构建具有语义规则的建筑物边缘信息;2)通过剖线距离相似性聚类以及BIP优化方法,为每一条边缘线提取一条对应的规则立面结构;3)最后,利用过程式建模方法,以建筑物边缘线为基础,以其立面结构为引导,恢复建筑物精细结构并重建LOD-2建筑物模型。更多还原

【Abstract】 The levels of detail of three-dimensional building models usually include four typical models from LOD-0 to LOD-3,with different geometric and semantic information for different levels of application requirements.In particular,the LOD-2 model that can distinguish from the roof and fa(?)ade structure of a building constitutes the skeleton structure of Smart City and has the most extensive application in urban planning and construction management.Due to the fa(?)ade visibility,intuitive expression and multi-view redundancy,the oblique photogrammetry has become the main data source for large-scale urban LOD-2reconstruction.The existing LOD-2 reconstruction methods based on oblique photogrammetry suffer from: 1)the technical bottleneck problems of difficult distinction and inaccurate extraction;2)topological reconstruction problems caused by data quality issues such as high noise,severe occlusion,and missing semantic information,resulting in low automation of LOD-2 model reconstruction and poor model quality.To overcome these,the paper makes full use of the available crowdsourcing contour features of building and intensively investigates the reconstruction at levels of detail of building models by means of contour features from oblique photogrammetry,with the aim of solving the problem of identification,extraction,and levels of detail model reconstruction for the buildings with multi-type and multi-shape.The specific research is as follows:(1)Multi-entity registration method for existing building contour features and oblique photogrammetric point cloudDifferences in the source and presentation between existing building contour feature data and oblique photogrammetric point cloud,inevitably lead to location bias.The accurate registration of the two is therefore extremely essential.The paper proposes a face-line-point multi-entity registration method for building contour features and oblique photogrammetric point clouds,using planar features as primitives to extract building fa(?)ade contour features from point clouds;using line features as primitives to construct line junction features,and adopting a RANSAC-based matching method to search for corresponding line junction features between building contour data and point cloud;using points as primitives,the transformation matrix is computed using a global least-squares optimization method for the intersections of line structure features with correspondence,achieving accurate registration of building contour data and point clouds.(2)Reliable plane extraction method of buildings via global normal refinement from noisy oblique photogrammetric point cloud.Models of urban buildings are often complex,but they can be assumed to be a combination of multiple simple planes,so it is fundamental to extract reliable and accurate planar primitives from building point cloud.However,duo to the problem of high noise in oblique photogrammetric point clouds,traditional regional growth method is inefficient in extracting complete planar primitives.Therefore,we propose a reliable plane extraction method of buildings via global normal refinement from noisy oblique photogrammetric point cloud.1)constructing supervoxels with strict planar features using similarity clustering;2)introducing the concept of maximum planar support regions and merging adjacent voxels to construct maximum planar support regions;3)adopting the point cloud local-global consistency spatial constraint method to transform local features into global features and optimize the point cloud normal;4)finally using the maximum planar support region guided planar extraction method to achieve accurate extraction of building planes.(3)Detailed structure optimization method for buildings guided by fa(?)ade structures of contour features.The finite parametric primitive models of existing model-driven approaches are difficult to support modeling complex environments,while data-driven approaches are constrained by point clouds noise,even the simplest plane primitives are difficult to be extracted fully and accurately,much less accurately recover the regular structure between the primitives.Therefore,we propose a detailed structure optimization method for buildings guided by fa(?)ade structures of contour features.1)constrained by the crowdsource building contour data,the building edge information with semantic rules is constructed based on the accurately extracted building vertical planes;2)generating a corresponding regular fa(?)ade structure for each edge by distance similarity clustering of the profiles and BIP optimization;3)finally,a procedural modeling approach is used to recovery the fine structure of the building and reconstruct the LOD-2 building model,using the building edge lines as basis and their fa(?)ade structures as guide.更多还原