| 摘要: | 一般大型物件，如客機、橋梁等之瑕疵檢測，因物件巨大又高，人類檢測必須依靠不同工具或爬高才能達成檢測目的。爬高危險性增加，靠其他工具費時，成本又高，效率也無法提升，因此本論文旨在建立一套基於無人機之大型物件外觀巡檢系統，能使無人機沿著物件外型以最短路徑自動飛行，並且控制雲台自動調整角度，以確保雲台相機的拍攝範圍能完整覆蓋待檢測物件，蒐集到物件的所有外觀影像後，再利用深度學習技術偵測物件外觀是否有瑕疵，協助人類更方便更有效率的完成檢測任務。
考量到室內與室外的通用性，此系統基於視覺之同步定位與地圖構建(Visual Simultaneous Localization and Mapping, V-SLAM)達成無人機的定位，因此不論GPS是否收訊良好皆能進行自動巡檢飛行。此系統可分為三個功能模組，第一部分為面對陌生的待檢測物件時，先透過手動控制無人機飛行，蒐集待檢測物件的360^°影像，便能進行運動恢復結構 (Structure from Motion, SfM) 建立物件的點雲模型，再設計點雲前處理演算法校正點雲模型的誤差。第二部分為透過點雲處理找出適合無人機自動飛行的導航點，以及在各個導航點中找出合適的無人機偏航(yaw)角度與雲台角度，再利用最佳化方法解決旅行推銷員問題(Travelling Salesman Problem, TSP) 找出路徑距離最短的導航點飛行順序，完成導航路徑規劃後，往後面對相同的物件即可直接進行自動巡檢飛行，不須再手動飛行建立點雲。第三部分為外觀瑕疵偵測，系統可偵測出物件外觀影像中的瑕疵，並且根據較常見的瑕疵種類進行分類，本論文考慮的瑕疵是以大型客機的瑕疵為對象，分別有掉漆、鏽蝕、凹痕三種類別。本論文使用OpenSfM建立物件點雲；利用隨機抽樣一致性演算法 (Random Sample Consensus, RANSAC) 與主成分分析 (Principal Component Analysis, PCA) 進行點雲前處理之校正；採用了基因演算法(Genetic Algorithm)計算導航點的飛行順序；使用Scaled-YOLOv4網路架構偵測與辨識物件外觀影像中的瑕疵。本論文在模擬環境與實體實驗中均驗證其可行性，在以汽車為檢測對象的實驗測試中，此系統的拍攝覆蓋率達100%，瑕疵偵測準確率達92.7%，成果顯示本論文提出的方法能拍攝到完整的物件表面影像且偵測出瑕疵。;In general, the defect inspection of large objects, such as airliners, bridges, etc., must rely on different tools or climb to high places to achieve the inspection because the objects are huge and high. However, climbing to high places is dangerous, and relying on other tools is time-consuming and costly. Therefore, this thesis aims to establish a system for detecting defects in the surface of large objects based on a drone. The system enables the drone to fly along the object′s exterior with the shortest path and adjust the angle of the gimbal such that the drone′s flying ensures complete coverage of the object by the gimbal camera. After the drone collects all the images of the object′s surface, it uses deep learning technology to detect any defects in the object′s surface. The system can help humans to complete the inspection task more easily and efficiently. To consider the usability of indoor and outdoor, the system uses Visual Simultaneous Localization and Mapping (V-SLAM) as the position control for the drone. The drone can fly stably regardless of the signal quality of the GPS. There are three function modules in the system. The first part is to collect 360 degree images of the object by manually controlling the drone when facing an unfamiliar object. Then, we use Structure from Motion (SfM) to build a point cloud of the object. Finally, the point cloud pre-processing algorithm is used to reduce the error of the point cloud. The second part is to find the appropriate navigation points for the drone to fly automatically and look for the suitable yaw angle and gimbal angle at each navigation point. Then we use the optimization method to solve the Travelling Salesman Problem (TSP) to find the shortest fly path. After completing the path planning, the same object can be flown directly for automatic inspection without needing to fly to create a point cloud manually. The third part is the surface defect detection. The system detects defects in the surface image of an object and classifies them according to the more common types of defects. The defects considered in this thesis are paint loss, corrosion, and dent on large airliners. Here, we use the OpenSfM to create point clouds of objects, use Random Sample Consensus (RANSAC), and Principal Component Analysis (PCA) in pre-processing to calibrate the point clouds. Then we use the genetic algorithm to compute flight sequences of navigation points and use the Scaled-YOLOv4 network architecture to detect and classify defects on the surface of objects. The feasibility of this thesis is demonstrated in both simulated and physical experiments. In an experimental test with a car as the target, the system achieved 100% coverage rate and 92.7% accuracy in detecting defects, showing that the proposed method can capture the complete object surface image and detect defects. |
