基於視覺的端對端自主無人機系統架構：應用於無 GPS 環境下之結構檢測

Abstract

在地震等天災發生後，城市基礎設施經常遭受嚴重破壞，並導致生活基本服務中斷。災後能否快速且可靠地完成結構檢測，關係到公共安全、建築物完整性評估與復原作業的效率。然而，傳統檢測方式不僅費時費力，還常伴隨人員進入危險現場的風險。雖然過去已有不少研究探討無人機（UAV）在導航與定位上的應用，然而針對災後結構檢測的完整作業流程仍相對缺乏。為了解決上述問題，本研究提出一套基於視覺的無人機自主結構檢測框架，專為 GPS 訊號缺乏的環境而設計。此系統採用端到端整合架構，從飛行前的路徑規劃，到飛行中利用「最佳視角搜尋（Find Best View, FBV）」演算法進行動態決策，再到飛行後的結構元件識別與損傷分類，全程自動化。我們以自組無人機為實驗載具，搭載 IntelRealSense T265 視覺慣性測程（VIO）相機，實現無 GPS 環境下的即時定位。系統在 Xu 等人所提出的導航架構基礎上，加入基於規則的 FBV 演算法，以便在飛行過程中根據結構特徵動態調整拍攝角度。影像採集後，先透過 You Only LookOnce (YOLO) 物件偵測模型辨識結構元件，再結合本研究團隊改良自國立臺灣大學張家銘教授團隊研究的情境感知分類模型，對影像進行損傷程度評估。實驗結果顯示，FBV 演算法可提升無人機在影像採集階段的自主判斷能力，使拍攝視角更具針對性；此外，將 DataCenterHub 資料集中之領域特定影像納入訓練，令模型在 IoU 0.5 條件下的平均精準度（mAP@0.5）達到 69.0%，最佳視角識別成功率（FBV-SS）為 68.05%，結構損壞指標識別成功率（SDI-SS）為 83.33%。本研究所提出之 UAV 自主檢測框架，不僅具備高效率與擴展性，更能自動蒐集視覺資料、辨識結構元件並分類損傷，於災後複雜的都市環境中，協助進行即時且具依據的決策。

Natural disasters such as earthquakes pose significant threats to urban infrastructure, often resulting in widespread damage and the disruption of essential services. In the aftermath of such events, rapid and reliable structural inspection is vital to ensure public safety, assess building integrity, and facilitate effective recovery operations. However, conventional inspection methods are typically labor-intensive, time-consuming, and pose risks to human inspectors. While previous research has explored UAV-based navigation and localization, comprehensive solutions for autonomous structural inspection remain limited.

To address these challenges, this study proposes a vision-based framework utilizing Unmanned Aerial Vehicles (UAV) for autonomous structural inspection in Global Positioning System (GPS)-denied environments. The proposed system constitutes a fully integrated, end-to-end solution encompassing the entire inspection workflow—from pre-flight path planning and mid-flight decision-making using a Find Best View (FBV) algorithm, to post-flight analysis involving structural component recognition and damage classification.

The framework is implemented on a custom-built UAV platform equipped with an Intel RealSense T265 visual-inertial odometry (VIO) camera to enable real-time localization in GPS-denied settings. Building upon the navigation framework introduced by Xu et al., this research incorporates a rule-based FBV algorithm to guide image acquisition porcess dynamically during flight. A You Only Look Once (YOLO)-based object detection model is employed to identify structural elements, followed by a context-aware classification model—adapted from work at National Taiwan University (Professor ChiaMing Chang), for post-flight damage assessment.

Experimental results demonstrate that the integration of the FBV algorithm enhances autonomous decision-making during image acquisition by adjusting the UAV’s viewpoint based on contextual structural cues. Additionally, augmenting the training dataset with domain-specific imagery from the DataCenterHub dataset significantly improved detection performance, achieving a mean Average Precision at IoU 0.5 (mAP@0.5) of 69.0%, Find-Best-View Success Score (FBV-SS) of 68.05%, and Structural Damage Index Success Score (SDI-SS) of 83.33%.

This research advances the field of post-disaster structural assessment by introducing a robust, efficient, and scalable UAV inspection framework capable of autonomous visual data collection, structural recognition, and damage classification, thereby supporting timely and informed decision-making in complex urban environments.