A Deep Learning Approach for Road Damage Detection Using UAV-Based Imagery

Abstract

Road networks are essential to contemporary transportation, and early detection of faults is crucial for ensuring safety and reducing maintenance expenses. Traditional road inspection techniques predominantly depend on manual surveys, considered time-consuming, labor-intensive, and frequently susceptible to human mistake. This study presents an automated approach for identifying road damage from photos obtained from Unmanned Aerial Vehicles (UAVs) utilizing deep learning methodologies. The system analyzes aerial photographs captured by drones and employs object detection techniques to recognize various road issues, including cracks, potholes, as well as surface degradation. Previous methodologies frequently employed models such as YOLOv5 for detecting tasks; however, these techniques may encounter difficulties in sustaining constant accuracy across varying illumination and ambient conditions. This study applies and assesses various deep learning models to enhance performance, utilizing YOLOv5 as the the beginning, YOLOv7 as the enhanced model, & YOLOv8 as the advanced version. The comprehensive system comprises multiple stages: dataset upload, preprocessing of data, dataset partitioning, model deployment, and performance evaluation using graphical analysis. The primary aim of this methodology is to improve detection precision and processing efficiency for the application of advanced YOLO structures, specifically optimized for UAV-based datasets. The experimental results demonstrate that the latest models attain superior accuracy and expedited detection relative to the basic model. The suggested system provides a dependable and adaptable solution for autonomous road condition monitoring, facilitating efficient maintenance of infrastructure and smart city redevelopment.