A modern technique to detect potholes by Computer Vision and Deep Learning

Abstract

Roads are connecting lines between different places and are used in our daily life but anomalies in road surface not only impact road quality but also affect driver safety, mechanic structure of the vehicles, and fuel consumption. Several approaches have been proposed to automatic monitoring of the road surface condition in order to assess road roughness and to detect potholes. Potholes are one of the main reasons behind the occurrence of road accidents. According to a report submitted by The Roads and Highways Department (RHD), around 25% roads of Bangladesh under the RHD across the country are in ”poor, bad or very bad” condition. This causes a lot of hassle and issues on the road for both humans and vehicles. Very often be cause of these potholes road accidents occur. Techniques for detecting potholes on road surfaces are being developed to provide real-time or offline vehicle control (for driver assistance or autonomous driving) as well as offline data collecting for road repair. For these reasons, researchers have looked into ways for detecting potholes on roads all over the world. This paper begins with a quick overview of the area before categorizing developed strategies into various groups. Then, by developing method ologies for automatic pothole detection, we present our contributions to the field. For this reason, we propose a deep learning approach that allows us to automatically identify the different kinds of road surface and to automatically distinguish potholes from destabilizations produced by speed bumps or driver actions. The system can detect potholes in different environments, lighting and weather conditions. We have trained and tested our model with a custom dataset which contains raw 3000 images with 1500 normal road images and 1500 images with potholes using deep learning algorithms. We have augmented these images and turned them into 120000 images so that the model can understand any image input in any scenario. In particular, we have analyzed and applied different deep learning models such as convolutional neural networks (CNN) and Yolov4. With these models we have achieved 97.35% accuracy with the CNN model and 87.6% accuracy with the YOLOv4 model.