本文为美国德州A&M大学科珀斯克里斯分校（作者：MICHAEL SCHWIND）的硕士论文，共129页。

运动结构（Structure from Motion，SfM）是一种摄影测量技术，通过重叠的二维（2D）图像序列来估计三维结构。它是在计算机视觉领域研究的，并在考古学、工程学和地球科学等领域得到应用。目前，许多SfM软件包允许生成三维点云。关于这些软件生成的地形数据在不同地形类型下的差异以及为什么它们可能产生不同的结果，目前几乎没有做过什么工作。

这项工作旨在比较和描述由三个不同的SfM软件包生成的点云之间的差异：两个著名的专有解决方案（Pix4D，Agisoft PhotoScan）和一个开源解决方案（OpenDroneMap）。利用DJI Phantom 3专业小型无人机系统（sUAS）对五种地形类型进行成像。这些地形类型包括沼泽环境、缓坡沙滩和防波堤、森林半岛、房屋和平坦的停车场。每一组图像都用软件进行处理，然后直接相互比较。在处理图像集之前，分析并选择软件设置，以便在三种软件类型中设置最相似的设置。这样做是为了尽量减少由不同设置引起的点云差异，然后对合成点云的特征进行了比较。此外，使用Riegl VZ-400扫描仪对平坦停车场进行了地面激光探测和测距（LiDAR）调查。以这些数据作为地面真实数据，以便对sUAS SfM点云进行精度评估。不同结果之间的差异不仅表现在云层的特征上，而且表现在精度上。这项研究使SfM摄影测量的用户能够更好地了解不同处理软件的比较以及SfM自动化在三维重建中的固有敏感性。由于本研究主要使用软件中的默认设置，因此研究参数变化对不同SfM软件包生成的点云数据集保真度的影响将有助于进一步的研究。

Structure from Motion (SfM) is a photogrammetric technique wherebythree-dimensional structures(3D)are estimated from overlappingtwo-dimensional(2D)image sequences. It is studied in the field of computer visionand utilized in fields such as archeology, engineering, and the geosciences.Currently, many SfM software packages exist that allow for the generation of 3Dpoint clouds. Little work has been done to show how topographic data generatedfrom these software differ over varying terrain types and why they mightproduce different results. This work aims to compare and characterize thedifferences between point clouds generated by three different SfM softwarepackages: two well-known proprietary solutions (Pix4D, Agisoft PhotoScan) andone open source solution (OpenDroneMap).Five terrain types were imagedutilizing a DJI Phantom 3 Professional small unmanned aircraft system (sUAS).These terrain types include a marsh environment, a gently sloped sandy beachand jetties, a forested peninsula, a house, and a flat parking lot. Each set ofimagery was processed with each software and then directly compared to eachother. Before processing the sets of imagery, the software settings wereanalyzed and chosen in a manner that allowed for the most similar settings tobe set across the three software types. This was done in an attempt to minimizepoint cloud differences caused by dissimilar settings. The characteristics ofthe resultant point clouds were then compared with each other. Furthermore, aterrestrial light detection and ranging (LiDAR)survey was conducted over theflat parking lot using a Riegl VZ-400 scanner. This data served as ground truthin order to conduct an accuracy assessment of the sUAS-SfM point clouds.Differences were found between the different results, apparent not only in thecharacteristics of the clouds, but also the accuracy. This study allows forusers of SfM photogrammetry to have a better understanding of how differentprocessing software compare and the inherent sensitivity of SfM automation in3D reconstruction. Because this study used mostly default settings within thesoftware, it would be beneficial for further research to investigate theeffects of changing parameters have on the fidelity of point cloud datasetsgenerated from different SfM software packages.

1.         引言

2.         文献回顾

3.         研究区域与数据集

4.         研究方法

5.         结果与讨论

6.         结论

更多精彩文章请关注公众号：