Deep Neural Networks for Asphalt Pavement Distress Detection and Condition Assessment

Monitoring the current state of critical infrastructure assets is a crucial task for asset managers to ensure structural integrity, operational safety, prevent deterioration, and optimize maintenance scheduling. Currently, the most widely used approaches for evaluating road pavement conditions rely on visual inspections conducted by specialized operators and rarely integrate ground-based technologies such as laser scanners, Falling Weight Deflectometers, and accelerometers. However, the high costs associated with maintenance operations and on-site surveys, which often require temporary partial or total closure of the infrastructure during testing, still limit the widespread adoption of these procedures at a network-scale level. In this context, recent advancements in Deep Learning methodologies and the broader field of Artificial Intelligence have enabled researchers to explore the potential of automated recognition and geospatial localization of damages affecting critical transportation assets, including road pavements on bridges, viaducts, and tunnels. This study presents an experimental application based on Deep Neural Networks (DNN) for the automatic detection and assessment of pavement distress. The methodology builds upon DNN algorithms incorporating object detection and semantic segmentation capabilities, including "YOLO v7" and "U-Net", respectively. Firstly, a dataset comprising top-down images of road damages that affect road pavements, obtained from various open-source datasets, was collected to train the object detection task of YOLOv7. This process enables the automatic identification of five distinct classes of pavement distress. Furthermore, a method of assessment of pavement conditions, based on the outcomes of the DNNs, was proposed to determine the level of pavement deterioration and tested on a local road investigated in Rome, Italy. The application of the proposed approach offers a reliable and promising methodology for the automated identification, geospatial localization, and severity assessment of pavement damages, enabling rapid and decisive actions by administrative authorities. The obtained information is crucial for prioritizing maintenance activities. The study's findings demonstrate the potential of DNNs and Deep Learning algorithms to complement Non-Destructive Remote Sensing technologies (e.g., NDTs, Laser Scanners, InSAR) by automatically localizing pavement and infrastructure damages. This study paves the way for integrated approaches in the next generation of Pavement Management Systems (PMS). © 2023 SPIE. All rights reserved.