Enhanced Pothole Detection in Road Condition Assessment Using YOLOv8

Abstract

Roads are vital for transportation, serving as a nation's lifelines. Maintaining road quality is paramount, and one key aspect is repairing potholes. Detecting potholes is challenging, especially in vast road networks like India's. Hence, there's a demand for high-speed, real-time pothole identification automation. Prior research has employed Convolutional Neural Networks (CNN) and various YOLO versions, such as YOLOv3 and YOLOv5, to address this issue. However, our study emphasizes the superiority of YOLOv8 in pothole detection. The purpose of the research is to train and assess the pothole detection ability of the YOLOv8 object detection algorithm. We evaluate accuracy, recall, and model size using a dedicated pothole dataset and compare it with other YOLO methods. Our tests show that pothole prediction is an effective feature for the YOLOv8 model with a low cost of computation. We obtain a small model size of 4.9 MB and an impressive Mean Average Precision (mAP) score of 78.27%. This study pioneers the application of the YOLOv8 algorithm in pothole detection, offering cost savings and faster identification for road maintenance a significant contribution to the field. Furthermore, our research establishes a seamless integration between pothole detection and actionable outcomes by facilitating the efficient transmission of identified pothole data to municipal corporations or governing bodies.