PROACTIVE DETECTION OF ALCOHOL IMPAIRMENT: LEVERAGING ARTIFICIAL INTELLIGENCE FOR ENHANCED TRAFFIC SAFETY

Drunk driving continues to seriously threaten public safety, affecting drivers, pedestrians, and other road users. Despite strict laws and the use of tools such as breathalyzers, the persistent prevalence of alcohol-impaired driving underscores the need for innovative prevention strategies to save traffic. Current detection systems are primarily reactive, with limited effectiveness in proactively identifying impairment, especially out-of-the-vehicle detection methods. To address this gap, this thesis introduces a novel approach that uses machine learning (ML) and deep learning (DL) techniques to classify people as intoxicated or sober using image data from outside the cars. The study uses a custom dataset that preprocesses and integrates images from the IMDB-Wiki and Drunk/Sober datasets, incorporating images with varying disruption and noise levels (ranging from 10% to 30%) to simulate different camera qualities outside the suspected person’s car. This noise and disruption addition enhances the dataset’s diversity to fit the real-life scenarios of collecting out-of-the-vehicle data from different sources with different capabilities. It ensures robustness in handling real-world conditions where the data can be collected from different sources, such as drones, closed-circuit television (CCTV) cameras, or autonomous cars driven near the suspected driver. We then train several ML and DL models to predict alcohol impairment by analyzing facial features and other image-based indicators. The results demonstrate significant advances in detection accuracy and reliability compared to existing systems, paving the way for proactive methods for preventing drunk driving that can affect traffic safety.