In our daily lives, when we want to connect with the outside world, we use roads as a connecting method between various locations. Road maintenance is a very important task to keep them safe. Potholes on roads may intensify the number of accidents, so detecting potholes and enlightening the concerned department can save the roads from getting worse. The traditional manual detection method was a time-consuming and laborious task. So In this paper, we have trained and tested different deep neural network-based models to detect the presence of potholes on roads. We have taken publicly available dataset from Kaggle and then potholes in road images are identified using object identification algorithms. Deep learning models for one-stage object detection that can be configured in a number of ways, including YOLOv4-CSPDarknet53, YOLOv5s, and modified YOLOv4-CSPDarknet 53 are used to compare their performance in detecting potholes. Performance is compared using Average precision, Average recall, mAP@50%, and inference time. The results showed that the modified YOLOv4 architecture outperformed compared with the other deep learning models with the highest Average precision (0.84), Average recall (0.79), F1 - Score (0.81), and mAP@50% (0.82).
Paper
Full text
Detection of Road Potholes Using Deep Learning Based Improved YOLOv4 Network
Semantic Scholar · Engineering · 2022
Abstract
In our daily lives, when we want to connect with the outside world, we use roads as a connecting method between various locations. Road maintenance is a very important task to keep them safe. Potholes on roads may intensify the number of accidents, so detecting potholes and enlightening the concerned department can save the roads from getting worse. The traditional manual detection method was a time-consuming and laborious task. So In this paper, we have trained and tested different deep neural network-based models to detect the presence of potholes on roads. We have taken publicly available dataset from Kaggle and then potholes in road images are identified using object identification algorithms. Deep learning models for one-stage object detection that can be configured in a number of ways, including YOLOv4-CSPDarknet53, YOLOv5s, and modified YOLOv4-CSPDarknet 53 are used to compare their performance in detecting potholes. Performance is compared using Average precision, Average recall, mAP@50%, and inference time. The results showed that the modified YOLOv4 architecture outperformed compared with the other deep learning models with the highest Average precision (0.84), Average recall (0.79), F1 - Score (0.81), and mAP@50% (0.82).