Enhancing Road Safety: Real-Time Surface Anomaly Detection Using YOLOv8

Abstract

Current navigation systems lack the capability to provide real-time information about road surface conditions, thereby posing significant safety risks. This study develops a robust system utilizing the YOLOv8 model for detecting road surface anomalies, including potholes, wet surfaces, sewer covers, drain holes, and unpaved roads. YOLOv8 demonstrates distinct advantages over its predecessors, such as enhanced accuracy, reduced inference times (18ms per image, 56 FPS), and an anchor-free detection mechanism, which simplifies training and improves detection precision. The study incorporates this model into a Flask web application, enabling real-time detection and visualization on a map. This integration, combined with persistent storage of detected anomalies along with their GPS coordinates, represents a novel approach to promoting road safety and transparency. Evaluations revealed a mean Average Precision (mAP) of 0.879 at Intersection over Union (IoU) 0.5 and 0.604 at IoU 0.95, validating YOLOv8's superior performance and efficiency.