Deep learning-based real-time pothole detection for avoiding road accident

Abstract

Bangladesh is a fast-developing country, and the number of roads increasing with it is immense. With the ever-growing amount of road comes the age-old problem of a pothole. This paper represents a model of deep learning-based, real-time pothole detection for finding and avoiding road accidents. Any types of image processingbased detection, in this case, pothole detection, are done through various steps. For example, collecting data sets is one of the most crucial steps to create any recognition system. Labeling an image means pinpointing the subject which we will be trying to find. Training the algorithm through those images to detect the subjects is critical in detecting potholes. In this research paper, to detect potholes from real-time videos, firstly, we collected data sets containing more than 600 images of potholes. After that, we labeled those images through labeling software. Then in chapter-1 we used those images to train the model (MobileNet, Inception-v3) which was detecting potholes from still photos given to it. Next, we used YOLOv5 to detect potholes from real-time feeds. In this proposed system, by using the real-time feed, potholes will be detected. Moreover, this will help the masses to detect potholes on roads to avoid accidents, and it will also help people related to the road works to find the potholes for further road maintenance.