A Semantic Segmentation Approach for Road Defect Detection and Quantification

Abstract

Automated visual detection and quantification of road defects has been a hot research topic for quite a long time due to its practical importance for road maintenance and traffic safety. However, uncertainties associated with the 2D images, such as non-uniformity of defects, insufficient background illumination, and etc., make it a challenging problem. This research work aims to solve the problem by employing a deep learning based approach. Specifically, image segmentation has been carried out, using a convolutional encoder-decoder model, to segment various defects from the non-defect area of the road. The method lead to a reasonable segmentation of different defects. Consequently, the extracted defect areas, in terms of number of pixels, are used to derive road condition indices being followed in Germany. In comparison, the indices derived using deep learning based approach are found to more accurate than those derived using conventional approach.