Distress Recognition in Unpaved Roads Using Unmanned Aerial Systems and Deep Learning Segmentation

Abstract

Effective condition assessment of road networks has been known to decrease road maintenance expenses and operation cost of the users. Several automated methods, such as computer vision–based systems, have been developed in this area, and the emergence of low-cost unmanned aerial systems (UAS) has encouraged UAS-based condition assessment of the road surfaces. The majority of the existing systems are developed for paved roads and there is limited research on vision-based assessment of unpaved roads. This paper introduces a framework to use deep neural networks and UAS to detect major distresses on unpaved road surfaces. The proposed method includes two parts: the first module segments the road surface pixels in UAS-captured frames, and the second module identifies distresses on the segmented road surface. Different deep neural network architectures were trained using transfer learning for the two-stage segmentation of the distresses. The results showed a promising performance in segmentation of road pixels, with more than 93.5% of intersection over union, and the defect classifier provided intersection over union rates of more than 86% in segmentation of potholes and washboardings.