Abstract: The performance of deep learning-based computer vision systems for road infrastructure assessment is hindered by the scarcity of real-world, high-volume public datasets. Current research predominantly focuses on crack detection and segmentation, without devising end-to-end systems capable of effectively evaluating the most affected roads and assessing the out-of-sample performance. To address these limitations, this study proposes a public dataset with annotations of 7099 images and 13 types of defects, not only based on cracks, for the confrontation and development of deep learning models. These images are used to train and compare YOLOv5 sub-models based on pure detection efficiency, and standard object detection metrics, to select the optimum architecture. A novel post-processing filtering mechanism is then designed, which reduces the false positive detections by 20.5%. Additionally, a pavement condition index (ASPDI) is engineered for deep learning-based models to identify areas in need for immediate maintenance. To facilitate decision-making by road administrations, a software application is created, which integrates the ASPDI, geotagged images, and detections. This tool has allowed to detect two road sections in critical need of repair. The refined architecture is validated on open datasets, achieving mean average precision scores of 0.563 and 0.570 for RDD2022 and CPRI, respectively

Authorship

Download reference