整合影像辨識與 SLAM 技術建立 UAV 橋梁劣化即時檢測系統

Abstract

橋梁作為交通運輸系統中關鍵的基礎建設，其結構安全攸關公眾生命財產之保障與基礎設施之永續運作。隨著橋梁使用年限逐漸提高，結構劣化問題日益嚴重，傳統仰賴人力進行的橋梁檢測方法，存在檢測效率低落、人力成本高昂、作業環境危險及檢測結果主觀性高等諸多限制，難以因應大量橋梁設施的日常維護需求。面對此一挑戰，如何發展具即時性、自動化、低成本與高準確率之橋梁檢測技術，已成為當前土木工程領域的重要研究議題。

為解決上述問題，本研究提出一套結合深度學習影像辨識與同步定位與地圖建構（Simultaneous Localization and Mapping, SLAM）之無人機（Unmanned Aerial Vehicle, UAV）智慧檢測系統。系統架構包含前端影像蒐集、中段即時推論與定位、以及後端三維建模與資料整合等模組。前端部分以無人機搭載光學相機進行橋梁影像拍攝，並以 NVIDIA Jetson Orin Nano 作為邊緣運算平台，即時執行影像辨識及同步定位與地圖建構。

影像辨識模組採用 YOLOv9 進行多類別劣化偵測，針對常見缺陷類型如裂縫、剝落與白華進行訓練與評估，訓練資料來自公開資料集，測試資料則為實地拍攝之影像，並透過灰階轉換與超參數調整優化模型性能。同步定位與地圖建構部分，採用 ORB-SLAM3 作為核心演算法，於飛行過程中估算無人機姿態並同步建立稀疏地圖。後續則結合其輸出的關鍵幀與相機外參資訊，匯入 COLMAP 進行稠密點雲重建。為實現缺陷資訊於三維空間中的準確標註與視覺化展示，本研究進一步整合建築資訊模型（Building Information Modeling, BIM），並透過點雲配準技術進行空間對齊，最終輸出包含缺陷位置、構件對應、TWD97 坐標與影像標註之結構化檢測報告。

實驗結果顯示，最終 YOLOv9 模型於本研究自建測試集上達到 mAP@0.5 為 0.657，於道南橋進行之現地飛行測試中，可穩定完成即時辨識與定位作業，成功標註 25 處劣化並建立完整三維點雲模型，Micro F1-score 達 0.656，驗證系統整合效能與即時性優勢。整體而言，本研究所提出之系統具備高度模組化、即時運作、低成本與高擴充性等特性，能有效協助橋梁管理單位進行結構檢測與缺陷紀錄，提升傳統檢測流程之智慧化與作業效率，並提供具實務應用潛力之技術解決方案。

Bridges are critical components of transportation infrastructure, and their structural integrity is essential for public safety and long-term serviceability. As aging structures face increasing deterioration, traditional manual inspections have proven inefficient, costly, hazardous, and prone to subjective errors—making them unsuitable for large-scale routine monitoring. Addressing these challenges requires an automated, real-time, and accurate inspection solution.

This study proposes a UAV-based intelligent inspection system integrating deep learning and Simultaneous Localization and Mapping (SLAM). The system consists of three modules: image acquisition via UAVs, real-time inference and localization using the NVIDIA Jetson Orin Nano, and 3D reconstruction and defect annotation. YOLOv9 is employed to detect common defects such as cracks, spalling, and efflorescence. The model is trained on public datasets and tested with field-collected images, with performance enhanced through grayscale preprocessing and hyperparameter tuning.

For localization, ORB-SLAM3 estimates the UAV’s pose and generates sparse maps in real time. Keyframes and camera poses are then used in COLMAP to reconstruct dense point clouds. These are aligned with Building Information Modeling (BIM) via point cloud registration to produce accurate 3D annotations and structured inspection reports containing defect locations, component IDs, and TWD97 coordinates.

Experimental results show that the optimized YOLOv9 model achieved an mAP@0.5 of 0.657. Field tests at Daonan Bridge demonstrated real-time detection and mapping of 25 defects with a micro F1-score of 0.656. The proposed system is modular, low-cost, and scalable, offering an effective solution for automating bridge inspection and enhancing maintenance workflows.