Deep Learning Framework For Intelligent Pavement Condition Rating: A direct classification approach for regional and local roads
Document Type Article
https://www.sciencedirect.com/science/article/pii/S0926580523002054
Waqar S. Qureshi, David Power, Ihsan Ullah, Brian Mulry, Kieran Feighan, Susan McKeever, Dympna O'Sullivan, Deep learning framework for intelligent pavement condition rating: A direct classification approach for regional and local roads, Automation in Construction, Volume 153, 2023.
https://doi.org/10.1016/j.autcon.2023.104945
Abstract
Transport authorities rely on pavement characteristics to determine a pavement condition rating index. However, manually computing ratings can be a tedious, subjective, time-consuming, and training-intensive process. This paper presents a deep-learning framework for automatically rating the condition of rural road pavements using digital images captured from a dashboard-mounted camera. The framework includes pavement segmentation, data cleaning, image cropping and resizing, and pavement condition rating classification. A dataset of images, captured from diverse roads in Ireland and rated by two expert raters using the pavement surface condition index (PSCI) scale, was created. Deep-learning models were developed to perform pavement segmentation and condition rating classification. The automated PSCI rating achieved an average Cohen Kappa score and F1-score of 0.9 and 0.85, respectively, across 1–10 rating classes on an independent test set. The incorporation of unique image augmentation during training enabled the models to exhibit increased robustness against variations in background and clutter.