本文为美国罗彻斯特理工学院（作者：Jie Zhang）的硕士论文，共116页。

随着具有中小型传感器的低成本数码相机越来越多，高分辨率机载图像的获取越来越容易，增大了为城市地区建立三维模型的可能性。资产评估或灾后恢复需要高精度的城市建筑模型。许多自动建模和重建方法与激光探测测距（LiDAR）数据一起应用于航空图像。如果没有提供激光雷达数据，就必须采用手动的人工步骤，这将导致半自动化技术。密集点云的自动提取可以辅助三维城市模型的自动提取。点云密度越大，建模越容易，精度越高。此外，倾斜的空中图像比低空图像（如建筑高度和纹理）能够提供更多的立面信息。因此，需要设计一种从倾斜图像中自动提取密集点云的方法。

本文提出并测试了一种改进的从倾斜图像中自动提取密集点云的工作流程。研究结果表明，改进后的工作流程运行良好，从两幅倾斜图像中提取出一个密度很高的点云，在平坦区域的精度略高于原工作流程提取的点云。最初的工作流程是由罗切斯特理工学院（RIT）以前的研究建立起来的，用于从nadir图像中提取点云。对于倾斜图像，首先在特征检测部分提出了一种改进方法，将尺度不变特征变换（SIFT）算法替换为仿射尺度不变特征变换（ASIFT）算法。然后，为了实现一个非常密集的点云，在第二次改进中实现了半全局匹配（SGM）算法，从立体图像对中计算出视差图，然后再将像素重新投影回点云。在第三次改进中增加了噪声消除步骤。

与原始工作流程的结果相比，修改后输出的点云密度更高。最后对改进后的工作流程中提取的点云进行了精度评估。从两个平面区域中，在原始和改进后的工作流节点上选择点的子集，然后分别将平面拟合到这些子集上，并比较了拟合平面上各点的均方误差。改进后的工作流中的点子集MSE比原始工作流略低。这表明，更密集、更精确的点云可以为屋顶提取清除边界轮廓，提高三维点云配准的特征检测可能性。

With the increasing availability oflow-cost digital cameras with small or medium sized sensors, more and moreairborne images are available with high resolution, which enhances thepossibility in establishing three dimensional models for urban areas. The highaccuracy of representation of buildings in urban areas is required for assetvaluation or disaster recovery. Many automatic methods for modeling andreconstruction are applied to aerial images together with Light Detection andRanging (LiDAR) data. If LiDAR data are not provided, manual steps must beapplied, which results in semi-automated technique. The automated extraction of3D urban models can be aided by the automatic extraction of dense point clouds.The more dense the point clouds, the easier the modeling and the higher theaccuracy. Also oblique aerial imagery provides more facade information thannadir images, such as building height and texture. So a method for automaticdense point cloud extraction from oblique images is desired.

In this thesis, a modified workflow for theautomated extraction of dense point clouds from oblique images is proposed andtested. The result reveals that this modified workflow works well and a verydense point cloud can be extracted from only two oblique images with slightlyhigher accuracy in flat areas than the one extracted by the original workflow.The original workflow was established by previous research at the RochesterInstitute of Technology (RIT) for point cloud extraction from nadir images. Foroblique images, a first modification is proposed in the feature detection partby replacing the Scale-Invariant Feature Transform (SIFT) algorithm with theAffine Scale-Invariant Feature Transform (ASIFT) algorithm. After that, in orderto realize a very dense point cloud, the Semi-Global Matching (SGM) algorithmis implemented in the second modification to compute the disparity map from astereo image pair, which can then be used to reproject pixels back to a pointcloud. A noise removal step is added in the third modification. The point cloudfrom the modified workflow is much denser compared to the result from theoriginal workflow. An accuracy assessment is made in the end to evaluate thepoint cloud extracted from the modified workflow. From the two flat areas,subsets of points are selected from both original and modified workflow, andthen planes are fitted to them, respectively. The Mean Squared Error (MSE) ofthe points to the fitted plane is compared. The point subsets from the modifiedworkflow have slightly lower MSEs than the ones from the original workflow,respectively. This suggests a much more dense and more accurate point cloud canlead to clear roof borders for roof extraction and improve the possibility of3D feature detection for 3D point cloud registration.

1 引言

1.1 三维建模与重构

1.2 点云提取

1.3 本文组织架构

2 研究目标

3 相关数据

3.1 象形图像

3.2 部分测试图像

3.3 象形图像 vs. 测试图像

4 设计方法

4.1 已有工作

4.2 第一次改进：仿射尺度不变特征变换（ASIFT）

4.3 第二次改进：半全局匹配（SGM）

4.4 第三次改进：噪声消除

5 研究结果

5.1 以前的研究结果

5.2 第一次改进的结果：ASIFT

5.3 第二次改进的结果：SGM

5.4 第三次改进的结果：噪声消除

5.5 精度评估

6 结论

7 未来工作展望

更多精彩文章请关注公众号：