| 摘要: | 本論文主要設計一架搭載 GPS 定位系統與單目鏡頭的無人航空載具 (UAV)，針對大型物件場景進行點雲建置與外觀瑕疵檢測。為了實現這一目標，透過點雲建置、導航點設計與路徑規劃、飛行控制與高度補償，提升了無人機於戶外巡檢任務中的穩定性，以及瑕疵偵測等多項技術整合，完成對目標物件的全面檢測，同時，該方法也降低了工作人員因為攀高操作伴隨的風險。
在點雲建置部分，無人機以 Althold 或 Loiter 模式進行手動繞物飛行，並儲存帶有 GPS Exif 資訊的 RGB 影像。後續使用 OpenSFM 進行點雲建模，建立大型物件場景。同時，提出改進地面法向量厚度分析法，結合 RANSAC 移除點雲異常值，並使用 DBSCAN 聚類篩選出目標物件。最後進行主成分分析 (PCA) 將目標物件轉換至理想導航座標。在導航點設計與路徑規劃方面，分析目標物件邊框產生導航點，解決導航點銳角交錯問題，並以群心演算法進行篩選。為了取得最佳化路徑，採用模擬退火演算法 (SA) 結合貪婪演算法 (Greedy Algorithm)，有效解決旅行商問題 (TSP)。此外，結合 ISS (Intrinsic Shape Signatures) 提取物件特徵點，確保無人機在繞物飛行時始終保持面向目標物件的穩定姿態。飛行控制與高度補償部分，則是使用 MobaXterm 上傳最佳路徑及飛行姿態至機載電腦，並使用 MAVROS 控制周邊設備。針對高度衰減問題，經過讀取氣壓計數據進行平滑化處理，分析高度變化趨勢，使用導航點上拉完成高度補償。此方法能分析不穩定氣壓計數據並加以判斷高度衰減趨勢。最後，針對繞物巡檢中所儲存的影像，採用 YOLOv8 網路架構進行瑕疵偵測，實驗結果顯示，大部分瑕疵檢測結果具有高信心度，能有效框選目標物件的外觀瑕疵，但部分視角仍存在瑕疵誤判情況。
;The main objective of this thesis is to design an unmanned aerial vehicle (UAV) equipped with a GPS system and a monocular camera for constructing point clouds of large object scenes. After capturing the point cloud data, the system filters and corrects the objects, generates navigation points, path planning and object defect detection. The thesis aims to address major problem including navigation points transformation, UAV pose designed and defining the UAV′s viewpoint. This research extends and improves upon existing UAV-based visual localization for large object detection, making it applicable to open spaces such as statues, cars ,rooftop or water tanks. The ultimate goal is to enhance performance in outdoor environments.
The contributions of this thesis include filtering the point cloud to identify primary target objects, designing navigation points, path planning and coordinate transformation of navigation points combined with the UAV′s viewpoint for object inspection. Using GPS combined with a barometer compensation to improved the UAV′s stability and flight endurance that during outdoor flights.
The experimental process is as follows: First, manual flight that using Althold or Loiter mode for circling. Point cloud scenes will built by processing RGB images with GPS EXIF data through OpenSfM. RANSAC is applied to process the ground data, and we propose a method to extend the ground thickness to reduce outliers. DBSCAN filters and cluster the target objects and PCA transforms them to an ideal navigation coordinate system. After obtaining the target object, navigation points are generated using the method proposed in this thesis. At the same time overlapping navigation points created by sharp angles are solved and filtered with centroid algorithm and the Traveling Salesman Problem (TSP) is solved by Simulated Annealing (SA) combined with a Greedy Algorithm. Additionally, Intrinsic Shape Signatures (ISS) are used to obtain the object′s shape features that define the drone′s viewpoint and orientation.After the path planning, the data is uploaded to the UAV′s offboard control computer, where MAVROS is used to control peripheral devices such as altitude trending analysis, height compensation, object circling inspections, and gimbal control. Finally, during the object circling inspection, images are processed using YOLOv8 for defect detection. |
