Road Defect Detection with Pothole Parameters Using Optimized YOLO Model

Road surface deterioration, particularly potholes, poses serious risks to traffic safety and vehicle integrity. Timely detection and localization are critical for effective maintenance and accident prevention. This paper presents a real-time road damage detection system focused on water-filled potholes, leveraging a deep learning approach based on the YOLOv7 object detection model. This work differentiates itself by addressing the specific challenge of detecting water-filled potholes-a task complicated by obscured visual cues-using a custom-annotated dataset tailored for this problem. The model was trained and optimized using the Tree-Structured Parzen Estimator (TPE) and deployed on Jetson AGX Orin platforms to validate its real-world feasibility. The optimized model achieved a mean Average Precision (mAP) of 0.7584 and an inference time of 16.7 ms, demonstrating the viability of an accurate, low-latency system in an edge computing environment. These findings demonstrate the feasibility of deploying an accurate, low-latency pothole detection system in edge computing environments. The proposed system has potential applications in smart transportation infrastructure, enabling proactive maintenance and enhanced road safety.