Monitoring the actual conditions of transport infrastructures is a priority for asset owners and administrators to ensure structural stability, operational safety and to prevent damages and deterioration. Currently, most protocols for assessing roads pavement conditions are based on visual on-site inspection conducted by specialized operators and, rarely, the local application of ground-based technologies and sensors such as terrestrial laser scanners (TLSs). However, the high costs of maintenance operations and on-site surveys still limit the application of these advanced procedures at the network level. Accordingly, the definition of innovative methodologies and procedures for continuous monitoring operations, especially for road pavements monitoring purposes, is still an open challenge. This research aims at investigating the viability of an automatic methodology for the detection and classification of pavement distresses based on Machine Learning (ML) models. More specifically, the methodology is based on the latest generation of Deep Neural Networks (DNN) algorithms, among which “YOLO v5”, and “Faster R-CNN”. To this purpose, an experimental evaluation is conducted by the acquisition and the processing of a publicly open-source dataset made available within the context of the “IEEE Global Road Damage Detection Challenge” (GRDDC2020). The implementation of the presented approach provides a technologically enhanced and reliable methodology for the provision of the identification and localization of roads damages to be more rapidly processed and conclusively actioned by asset owners and management agencies giving crucial information that could be implemented for the prioritisation of maintenance activities within Pavements Management Systems (PMSs). The outcomes of this study demonstrate that ML approaches and DNN algorithms, can be applied to complement Non-Destructive Remote Sensing technologies (e.g., ground-penetrating radars, Laser Scanners, satellite radar interferometry), localizing automatically the pavement damages, thereby paving the way for integrated approaches in the smart monitoring of infrastructure assets.
Menghini, L., Bella, F., Sansonetti, G., & Gagliardi, V. (2021). Evaluation of road pavement conditions by Deep Neural Networks (DNN): an experimental application. In Proceedings of SPIE - Earth Resources and Environmental Remote Sensing/GIS Applications XII (pp.21-30). Schulz, Karsten [10.1117/12.2599894].