Detection of pothole for repair works of asphalt flexible pavement optimization using YOLO

Abstract

In Malaysia, the government has consistently allocated a significant amount of budget in improving and maintaining the country's road service quality. To ensure best-value maintenance works, it is essential to carry out regular pavement distress surveys for efficient pavement management and planning. Pothole is a long-standing pavement distress issue and most highlighted pavement distress with many lives have been claimed over the years. A road surveyor detects the distress by manually inspecting the road surface, therefore it is a tedious, time-consuming, laborious and hazardous process. This study proposes a new method to assist road surveyors in detecting potholes using a deep learning approach. Road surface images with potholes are captured from the federal roads in Kedah. The LabelImg app is used as a tool to label the potholes in the road images by defining a bounding box around each pothole. Finally, the labelled pavement images are presented to a deep learning method called You Only Look Once (YOLO), to train the network in detecting the potholes from the given images. The YOLO was found to achieve promising results in detecting the potholes images under various conditions, with an accuracy of up to 80%. The area of the YOLO's bounding box was also beneficial in estimating the pothole volume for efficient pothole repair work.