Towards Safer Roads: Vehicle Detection, Counting and Classification in Challenging Foggy Conditions

Abstract

Monitoring highway traffic is an important task for any country since it involves the transportation of people and goods across geographical boundaries. Manual monitoring of vehicles travelling via highways was frequent in the past, but thanks to technological improvements, a number of studies and activities have been carried out to improve travel safety. Previous research has significantly contributed to our understanding of foggy conditions in traffic monitoring. For instance, the DAWN dataset provided insights with 1000 images of real-world fog, shedding light on the challenges of natural foggy scenes. Additionally, the work ’Semantic Understanding of Foggy Scenes with Purely Synthetic Data’ created a synthetic dataset by replicating Zurich in a virtual environment and adding synthetic fog for realism. The BDDIW dataset, an extension of BDD100K, explored fog’s impact on urban highways with images depicting various foggy scenarios. However, these works have limitations, including small dataset sizes, artificial scenes, and a limited representation of fog variations in public datasets. In this study, we assess the performance of the state-of-the-art YOLOv8 algorithm in monitoring highway traffic. Our focus encompasses vehicle detection, counting, and classification to comprehensively enhance traffic management. Additionally, we recognize the significance of addressing adverse weather conditions, particularly fog, which significantly increases the risk of road traffic accidents, especially during winter. To overcome this challenge, we created a dataset of 8,000 highway traffic images with varying levels of artificial fog: low, medium, and high. This dataset includes both the original, fog-free images and those with simulated fog. We used these foggy images to train the YOLOv8 algorithm, achieving an impressive mean average precision (mAP) score of 0.942. This result clearly demonstrates the algorithm’s exceptional ability to 1 List of Tables identify and categorize vehicles, even in challenging foggy conditions. In essence, this work evaluates YOLOv8’s performance on foggy images, yielding promising results. By harnessing advanced computer vision algorithms and proactively addressing the challenges presented by adverse weather conditions, this study makes a substantial contribution to improving highway travel safety and efficiency through the creation of a comprehensive dataset. The insights and discoveries presented herein provide invaluable guidance for traffic management authorities and policymakers, equipping them with the tools to enact highly effective strategies for mitigating the risks associated with foggy weather conditions.