Machine Learning Techniques for Road Health Monitoring

Asset owners and administrators have to monitor the current state of transportation infrastructures in order to guarantee structural stability, operating safety, and the prevention of damages and deterioration. Nowadays, most procedures for reporting road pavement conditions rely on visual inspections performed by experienced operators and, on rare occasions, the local deployment of ground-based technology and sensors such as terrestrial laser scanners. However, the high cost of on-site survey maintenance prevents the widespread use of such sophisticated methodologies. Consequently, the definition of novel procedures for continuous monitoring activities, especially road pavement conditions, remains an open problem. The goal of our research activities is to verify if a fully automated system for detecting and classifying pavement distresses can be devised and realized using Machine Learning (ML) models and low-cost equipment like a simple camera and off-the-shelf instruments. The proposed approach takes advantage of some of the most recent Deep Neural Networks (DNNs), including Faster R-CNN [3], EfficientDet [2], and You Only Look Once (YOLO v5) [1]. In order to evaluate the feasibility of our approach, we carried out extensive and in-depth experimental trials based on both ad-hoc collected data and a publicly available dataset released during the Global Road Damage Detection Challenge (GRDDC)1, a Big Data Cup organized as a part of the 2020 IEEE International Conference on Big Data (IEEE BigData 2020). Our experimental findings show that Machine Learning techniques, specifically DNN algorithms, can be used to complement non-destructive remote sensing technologies (e.g., laser scanners, ground-penetrating radars, and satellite radar interferometry) by automatically detecting and classifying road pavement distresses.