Advancing UAV-based inspection system : a novel AI-powered technique for crack detection, localization and quantification

Abstract

In the field of crack inspection for complex infrastructures, traditional methods have predominantly relied on costly structural health monitoring instruments and labor-intensive processes. However, the advent of Unmanned Aerial Vehicle (UAV) technology has introduced innovative and efficient solutions for inspecting large-scale infrastructure. Alongside the rapid advancements in UAV technology, the integration of artificial intelligence (AI) for object detection, classification, and segmentation has further demonstrated remarkable robustness and efficiency. This thesis presents an autonomous inspection system that leverages enhanced machine learning models on UAVs to enable crack detection, localization, and quantification throughout the structural inspection process. In exploring AI-based methods for autonomous infrastructure inspection, this research focuses on the precise detection and localization of defects within target structures. The proposed crack detection model, an enhanced version of the YOLOv4 object detection framework integrated with an attention module, exhibits real-time multi-crack detection capabilities. It accurately identifies crack positions within a global coordinate system. The system achieves a mean Average Precision (mAP) of 90.02%, surpassing the original YOLOv4 by 5.23%. Notably, the inspection system operates independently of predefined navigation trajectories, with experimental results underscoring its robustness and effectiveness in real-world crack detection and localization tasks. To further enhance the autonomous inspection system, this study introduces a novel crack quantification approach utilizing LiDAR scanning to create a 3D model of the target structure. A customized path planner ensures complete coverage of all crack points during inspection. Extensive testing of the proposed deep crack segmentation model, U-Net with Spectral Block and Self-Attention Module (USSA-Net), highlights its performance in crack quantification, achieving root-mean-square errors of 9.9mm and 6.2mm in horizontal and vertical crack thickness estimations, respectively. This research provides a comprehensive demonstration of UAV-based inspection systems for industrial applications, including defect detection, localization, and quantification. By integrating AI-powered components, it significantly advances the role of UAVs in infrastructure inspection, offering valuable insights for practitioners in structural health monitoring. Extensive simulations and flight tests validate the system's effectiveness, bringing the vision of intelligent inspection systems closer to reality.