桥梁在我国公路交通运输中扮演着重要作用，近二十年来桥梁数量的快速增长在促进交通发展的同时，也带来了更大的运维和养护压力。为了满足日益增长的桥梁定期巡检需求，基于无人机等平台的智能化表观病害采集技术受到关注，凭借其高效、高精度、自动化的优势具有广阔的运用前景。然而，由于相关研究处于探索阶段，基于无人机和机器视觉的智能检测方法缺乏完整的技术路线，对病害采集过程也缺乏定量化的研究。本论文主要基于无人机与机器视觉的桥梁表观病害智能检测方法，建构适用于桥检的软件和硬件系统，并对主要数据采集关键参数对处理效果的影响进行了定量研究，据此提出了完整的无人机桥检技术路线，取得的主要研究成果如下： （1）在系统建构部分，论文基于现有无人机平台和镜头建构了包含三种无人机、三种采集传感器等在内的硬件系统，基于YOLO v5算法和已有裂缝数据库建立了图像处理及病害识别软件系统，基于 SfM-MVS 算法建立了三维建模软件系统，搭建了一套完整的桥检机制。 （2）在现场试验部分，论文主要针对数据采集阶段拍摄距离、重叠率和光照条件这三个最为重要的参数进行定量研究，研究发现，重叠率对建模精度的影响最大，较低的重叠率将显著降低稀疏建模和稠密建模环节的精度，过低的重叠率甚至会导致建模失败；拍摄距离直接影响图像的影像分辨率，在稀疏点云重建阶段对特征点的匹配精度影响较小，但更近的拍摄距离可以提升稠密点云重建阶段的点云质量和病害识别精度；光照条件在非极端条件下对数据处理效果影响较小，但应避免光照过强或过弱条件下进行采集导致局部或整体建模失败的情况。 （3）在技术路线部分，论文基于第二章、第三章的研究成果，结合其他相关研究和工程中积累的经验，提出了包括硬件选取、方案设计、现场采集、后处理阶段的无人机桥检技术路线。同时，基于第三章试验结果，论文提出了对重叠率、采集距离、相机参数等外业工作中相关参数的选取建议。 本论文获得国家自然科学基金项目（52192662）的资助。
Bridge is a significant part in the whole highway transportation system. In the recent decades, there has been a huge increase in the number of highway bridges in China, which has not only promoted development in the transportation industry, but also simultaneously generated higher maintenance pressures. To address the growing demand for periodic bridge inspections, intelligent detection methods for bridge surface defects have emerged, mainly based on platforms such as unmanned aerial vehicles (UAVs). These methods have good application prospects because of their good performance in detection efficiency, precision, and automation. Nevertheless, due to the early stage of relevant research, intelligent detection methods based on UAVs and machine vision are in lack of systematic framework. Quantitative studies on the subject are also insufficient. This research establishes hardware and software systems based on UAVs and machine vision. Furthermore, in regard to the data acquisition stage, several field tests have been held to quantitatively investigate the effects of key factors. Last of all, a comprehensive UAV-based bridge inspection technical framework is proposed. Main results of the research are as follows:(1) In the section of inspection system, the research establishes a hardware system involving UAV platforms, collection sensors and other associated components. Meanwhile, it establishes image-processing and damage-identification software system based on YOLO v5, alongside a 3D-reconstruction system based on SfM-MVS algorithm.(2) In the section of field tests, quantitative analysis is held to investigate three key parameters in data acquisition: shooting distance, overlap rate, and illumination. Relevant result indicates that overlap rate has the most significant influence on the accuracy of 3D reconstruction, where lower overlap rates considerably reduce precision in both sparse and dense point cloud reconstruction phases, even potentially causing failure. Distance directly affects image resolution. It demonstrates minimal impact on sparse point cloud reconstruction, but at dense point cloud reconstruction stage lower distance leads to better point cloud quality and damage identification accuracy. Illumination shows relatively minor effects on data processing under non-extreme scenarios, though excessive or insufficient illumination should be avoided in data acquisition.(3) In the section of technical framework, in conclusion of findings from preceding chapters, the research proposes a UAV bridge inspection workflow including hardware selection, scheme design, data acquisition and processing stage. It further provides recommended range of key parameters in data acquisition based on the results of chapter three. This dissertation is subsidized by the National Natural Science Foundation of China (52192662).
