Detection of Asphalt Pavement Cracks with YOLO Architectures from Unmanned Aerial Vehicle Images

Monitoring road condition has been a strategic area of research in maintaining an extensive transportation infrastructure network. Although damages on road surfaces initially appear as slight cracks, the depth and danger of these damages may increase over time and changing weather conditions. Cracks on the road surface are one of the main factors affecting the performance of the road. Automatic detection of road cracks is an important task in road maintenance. However, automatic crack detection is a challenging application area due to the inhomogeneity of the density of cracks and the complexity of the background (e.g. low contrast with the surrounding coating and possible shadows of similar intensity). Recently, deep learning-based object detection and segmentation methods have begun to be used effectively in detecting cracks on road surfaces. In this study, a comparative analysis was carried out for the detection of cracks on road surfaces using the current versions of You Only Look Once (YOLO), a popular single-step object detection algorithm. The open source dataset UAPD, consisting of unmanned aerial vehicle (UAV) images, was used in the analysis. In the application carried out to detect different types of cracks with YOLOv5x, 0.639 mean average precision (mAP) and 0.759 sensitivity metrics were obtained. Using the YOLOv5x algorithm, the highest accuracy was achieved compared to other algorithms.