Importance of Contextual Information for the Detection of Road Damages

Abstract

In this paper we give an overview on methods for the optical detection of road surface damages and analyze the importance of contextual information. The objective is to improve the optical detection of road damages, especially potholes, based on images from windscreen mounted monocular cameras, as well as to reduce the complexity and thus save computing capacity. In order to achieve this, parts of the image that are not classified as road are preprocessed. Thus two different parts were implemented and analyzed: (i) a semantic segmentation for classifying the contextual information and (ii) a pothole detection.First, different semantic segmentation networks are compared by varying the training dataset and the number of predicted classes to find a trade-off between accuracy and the level of detail of the provided information. In the second step, this segmentation into road and other is used to evaluate the importance of the contextual information in the detection part. The detection accuracy is compared using varying input datasets by hiding and adjusting the non road parts. From the state of the art it can be derived that the elimination of contextual information reduces the complexity for the detection of road surface damages like potholes especially for computer vision (CV) based approaches. In addition, our results show that deep learning (DL) approaches need at least a simplified contextual information.