本文为美国罗切斯特理工学院（作者：Ming Li）的硕士论文，共101页。

高分辨率航空图像的日益普及，提高了城市场景的建模精度。这种高精度建模为灾难恢复和资产评估提供了强有力的参考。本文基于已有的机载斜向图像重建技术，利用多视场信息的优势，对建筑物模型进行重建。

本文在前人工作的基础上，提出了两种基于斜向图像点云的建模方案。在假定建筑物为立方体的情况下，第一种方案包括三个不同的步骤。平面估计旨在识别主要表面；边缘提取有助于检测并简化每个已识别表面的平面内边缘；模型构造将表面和边缘组装在一起，以普遍可接受的格式完成模型生成。我们发现，这种方案对于覆盖整个建筑四周的完整点云很有效。针对点云不能覆盖建筑物四周时的复杂情况，提出了第二种方法。利用主要平面上的最小包围盒估计建筑物的主要结构，其余的估计平面随后被连接到主结构，这样可以产生一个完美的建筑模型。

这些方案在来自多个源的点云数据集上进行了测试，包括图像衍生点云和激光雷达衍生点云。基于曲面的方法和基于最小包围盒的方法都显示了重建模型的能力，但两者都存在缺点。本文还对点云密度、拟合精度以及未来工作包括提高效率和鲁棒性等方面的局限性进行了讨论。

The increasing availability of high resolution airborne imageryincreases the accuracy of building modelling of urban scenes. This highaccuracy of building modeling offers a strong reference for disaster recoveryand asset evaluation. With the advantage of having more facade information,this thesis builds on previous efforts in building reconstruction from airborneoblique imagery. Based on previous work, this thesis presents two schemes toconstruct building models from point clouds derived from oblique imagery. Withthe assumption that buildings are in a cubic-shape, the first scheme consistsof three different steps. Plane estimation aims at identifying dominantsurfaces; edge extraction helps in detecting and simplifying in-plane edges ineach identified surfaces; model construction finishes the job of assembling thesurfaces and edges together and producing a model in a universally acceptedformat. We find this scheme works well with complete point clouds that coverall sides of the building. A second method is proposed to handle thecomplications when the point clouds do not cover all sides of the building. Themain structure of the building is estimated using minimum bounding box on the dominantplanes. The rest of the estimated planes are then attached to the main structure.The process can produce a water-tight building model. The schemes are tested onpoint cloud data sets from multiple sources, including both image derived andlidar derived point clouds. The surface based approach and minimum bounding boxbased approach both show the capability of reconstructing models, while both ofthem have disadvantages. The limitations such as density of point clouds;fitting accuracy; and future work, including increasing efficiency and robustness,are also discussed.

1 引言

2 项目背景

3 已有研究工作回顾

4 实验数据

5 基于单表面的方法

6 基于最小包围盒的方法

7 结果与讨论

8 结论与未来工作展望

下载英文原文地址：

http://page5.dfpan.com/fs/5lcc7jb232313249169/

更多精彩文章请关注微信号：