Distracted Driver Detection Using Advanced Deep Learning Models

Distracted driving is still a major source of road accidents around the world. This requires innovative and efficient solutions capable of real-time detection. Traditional monitoring systems often fail to capture subtle distractions. This paper presents a revolutionary ensemble deep learning framework that combines cutting-edge architectures with advanced optimization techniques for accurate, real-time distracted driver detection across diverse operational conditions. We created a weighted ensemble by integrating ResNet-101, EfficientNet-B7, and Vision Transformer architectures with proprietary attention methods and temporal consistency modules. The State Farm dataset was used in the evaluation, which covered ten distraction categories. Results show a 99.81% classification accuracy and real-time inference speeds of up to 70 FPS, exceeding existing approaches. Insights into feature importance and model scalability are also provided, offering a pathway for real-world implementation.