Transformer-CNN Hybrid Model for Real-Time Driver Drowsiness Detection

Driver drowsiness is a leading factor in global accident rates, sometimes resulting in significant injuries or even fatalities. Traditional fatigue detection systems depend on behavioural signals, physiological indicators, or vehicle movements. However, these technologies usually yield high false alarm rates and may not always perform in real time. Recently, deep learning models have demonstrated significant accuracy in detecting tiredness just through facial cues. This paper presents a unique transformer-CNN hybrid model that considers both Convolutional Neural Networks (CNNs) for capturing spatial features and Transformers for handling temporal sequences. The model is designed to efficiently process real-time video streams, ensuring it captures both small-scale facial changes and broader eye movement patterns. To evaluate its effectiveness, we compare our model against existing deep learning techniques using well-known datasets like NTHU-DDD and UTA-RLDD. Key performance indicators—accuracy, F1-score, and AUC-ROC—are analysed to see how well the model performs. Our Transformer-CNN hybrid outperforms conventional CNN and RNNbased models, offering better robustness across different lighting conditions and driver behaviours.