Evaluation of CNN-Based Approaches to Adverse Weather Image Classification for Autonomous Driving Systems

Weather image classification is a critical component of the vision systems in autonomous driving systems (ADSs), facilitating accurate decision-making across diverse driving conditions. Adverse weather conditions (AWCs) can significantly impair sensor data quality, diminishing the ADSs’ ability to interpret the surrounding environment. It is, therefore, essential for ADSs to effectively perceive and adapt to AWCs, ensuring enhanced performance and safety. This paper introduces a novel evaluation methodology for classifying AWC images using Convolutional Neural Network (CNN) models, with the goal of assessing their effectiveness for use in ADSs. The methodology provides a structured process for evaluating CNN models, taking into account key factors such as architectural designs, model sizes, diverse datasets, AWC scenarios, and real-time performance. A bespoke design framework is developed to guide the experimental modelling work, incorporating a range of representative CNN-based classification approaches and a variety of AWCs datasets and weather scenarios. This is followed by a comprehensive comparative performance analysis for both single-label and multi-label classification of AWCs images, which is grounded in an extensive experimental modelling effort and serves the purpose of validating the proposed novel evaluation methodology. The analysis systematically evaluates the performance of the targeted CNN approaches under consistent conditions, utilizing the same datasets and weather scenarios to provide a thorough and reliable comparison. Additionally, it includes performance testing on a small-scale embedded computing platform to examine real-time applicability. The findings and insights from this study aim to help researchers identify the most suitable CNN-based weather image classification approaches for their ADS application, ensuring alignment with their performance and operational requirements.