Unpaved roads’ surface evaluation using unmanned aerial vehicle and deep learning segmentation

Abstract

Unpaved roads have a significant role in Canada's transportation and service activities, accounting for close to 60% of Canada's total public road networks. Furthermore, they connect agricultural, mining, recreational areas, and small communities to the nearby towns and businesses. An effective maintenance program for a network of unpaved roads requires a detailed assessment of the road surface’s condition, and such assessment is usually made by visual inspections which can be time-consuming and error prone. The main part of these evaluations aims to identify distresses on the road surface, such as washboarding (corrugation), potholes, and rutting. Many research studies have developed methods to automate condition assessment of asphalt roads by combining machine learning algorithms and low-cost unmanned aerial vehicles (UAV), but the research on the automated assessment of unpaved roads is very limited. A system has been developed in this study to automate the assessment of unpaved roads by coupling computer vision methods, namely deep convolutional neural networks, and UAV-based imaging. This automated system could be used as an alternative method to reduce the need for human effort and possible manual errors, and therefore improve road maintenance programs in remote areas. The performance of the proposed method was evaluated using different test settings, and despite having some challenges, such as false positives, it showed promising outcomes that can contribute to the proposed purpose of this research. This proposed method has a potential for further improvement and the findings can be used as a basis for similar studies.