Deep Learning Approaches for Effective Fog Detection

ABSTRACT One of the primary challenges in ensuring road safety is effectively alerting drivers to adverse weather conditions, such as fog, which can severely impair visibility and increase the risk of accidents. Timely and accurate fog detection is crucial for providing drivers with the necessary warnings to adapt their driving behaviour and enhance safety. This paper presents a system designed to detect foggy scenarios and classify visibility levels, thereby enabling timely alerts for drivers to minimise the risks associated with reduced visibility. To achieve this, we have developed two new image datasets of road fog scenarios – Foggy‐Ceit 2023 and an extension to the Foggy CityScapes – DBF dataset – featuring both real and synthetic fog. Additionally, we compare various algorithms developed using classical vision techniques and deep learning methods (vision transformers [ViT] and EfficientNet). Finally, eXplainable artificial intelligence techniques are utilised to provide visual explanations and evaluate the performance of these models.