Heavy construction infrastructure including bridges, dams, high-rise buildings, and industrial facilities requires regularstructural health monitoring to ensure safety and prevent catastrophic failures. Traditional manual inspection methods arelabor-intensive, time-consuming, and often require scaffolding or specialized access equipment that puts inspectors at risk.This paper presents a novel drone-based crack analysis system implemented as a Python Windows application that processesreal-time video feeds and sensor data from unmanned aerial vehicles (UAVs) to automatically detect, classify, and analyzecracks in heavy construction structures. The system integrates a custom-trained YOLOv8 deep learning model using acomprehensive dataset of 25,000 annotated images from Roboflow, encompassing diverse crack types including hairlinecracks, structural cracks, surface cracks, fatigue cracks, and thermal cracks across various construction materials (concrete,steel, masonry). The architecture comprises four integrated modules: (1) Drone Communication Module that interfaces withDJI and custom UAV platforms via MAVLink protocol, capturing synchronized 4K video at 30fps and telemetry dataincluding GPS coordinates, altitude, orientation, and timestamps; (2) Real-Time Video Processing Pipeline implementingframe extraction, enhancement through adaptive histogram equalization and noise reduction filters, and parallel processingusing GPU acceleration; (3) YOLO-Based Crack Detection Engine featuring YOLOv8m architecture with 3.2 millionparameters, achieving mean Average Precision (mAP) of 0.943 at 0.5 IoU threshold and inference speed of 45 frames persecond on NVIDIA RTX 3060 GPU; (4) Crack Analysis and Reporting Module that quantifies crack dimensions (length,width, depth estimation), classifies severity into four levels (minor, moderate, severe, critical), analyzes potential causesbased on crack patterns (structural stress, thermal expansion, material fatigue, settlement), and generates comprehensiveinspection reports with geotagged crack locations. The model was trained on Roboflow datasets comprising 25,000 imageswith 45,000 annotated crack instances, augmented with synthetic variations (rotation, scaling, brightness adjustment,Gaussian noise) to improve generalization. Experimental evaluation on 50 real-world construction sites demonstratesdetection accuracy of 96.8%, precision of 95.2%, recall of 94.7%, and F1-score of 0.949. The system successfully identifiescracks as small as 0.5mm width from drone altitudes up to 15 meters. Sensor data fusion incorporating infrared thermalimaging detects subsurface cracks invisible to optical cameras, improving early detection by 34%. The Windows applicationprovides an intuitive GUI with real-time visualization, historical data comparison, automated report generation in PDF/Excelformats, and cloud synchronization for collaborative analysis. This work represents the first integrated drone-based crackanalysis system combining state-of-the-art YOLO deep learning with multi-sensor data fusion for comprehensive heavyconstruction structural health monitoring.
Mohd et al. (Mon,) studied this question.