Automated Segmentation and Deterioration Determination of Road Markings

Abstract

Road markings are used on pavement surfaces to provide guidance and information for drivers and pedestrians. When navigating road networks, awareness of upcoming hazards and critical information is essential for safe and comfortable driving. However, once the road markings get damaged from traffic application and environmental conditions, it becomes less efficient due to the removal of surface markings. Hence, it is crucial to determine road marking deterioration and propose appropriate rehabilitation methods. With the advancement of deep learning (DL) methods, as a branch of machine learning (ML), there is high potential for developing an automatic road marking detection algorithm with high accuracy and significantly shorter computational and analysis time. In this study, an automated algorithm for segmenting the road markings using mask region-based convolutional neural networks (mask R-CNNs) and determining their deterioration by image processing using the Otsu algorithm was developed. The developed mask R-CNN model used 6,500 and 3,500 images for training and validation, respectively. As a result, the mask R-CNN model could detect and segment road markings with an average accuracy of 97.10% for road marking object detection and 91.0% for road marking segmentation. Furthermore, the image processing algorithm obtained high precision of 92.0%. Therefore, the proposed method was found to be a promising approach to detecting and segmenting road markings together with determining their severity. Road markings play a vital role in traffic safety, so their evaluation and immediate rehabilitation are necessary. Recently, studies determining road marking defects have been conducted; however, they are time-consuming and require a retroreflectivity device. These methods, moreover, can only be applied for image classification and not damage evaluation, especially in extreme conditions. This research provides an automated process of determining the deterioration of road markings with high accuracy, addressing the previous issues discussed. In this study, segmenting road markings as new areas using mask R-CNN under extreme environmental and severely damaged conditions is presented. Using the segmented road marking, the damaged area whose pixel value is less than the threshold is calculated and compared with the whole area. The Otsu algorithm then was used to automatically collect the threshold value through the road marking image histogram. The proposed method can be used in different road networks, as applied to a 502-km road in Seoul. This study has great potential to replace manual visual assessments conducted by humans and retroreflectivity devices with higher precision and reliability. Furthermore, this study can aid pavement agencies in easily setting the deterioration criteria of road markings suitable for self-conditions.