Temporal Convolutional Network Approach for Video-Based Driver Drowsiness Detection

Abstract

Driver drowsiness remains a critical factor in road accidents worldwide, necessitating robust real-time detection systems for enhanced road safety. This research proposes a novel approach for drowsy driver detection by leveraging Temporal Convolutional Networks (TCN) to analyse sequential video frames and capture temporal dependencies in driver behaviour. Traditional computer vision methods often rely on static frame analysis, missing crucial temporal patterns indicative of drowsiness progression. Our methodology extracts spatiotemporal features from video sequences using a TCN architecture, enabling continuous monitoring of eye closure duration, blink frequency, head pose variation, and yawning patterns over time. The system employs preprocessing techniques including keyframe selection, face alignment and landmark detection analysis to ensure robustness under varying illumination and driving conditions. Experimental evaluation on benchmark datasets (NTHU-DDD) demonstrates that TCN approach achieves 96.7% detection accuracy with a false positive rate of 3.2%, outperforming conventional CNN and LSTM-based methods by 4.3% and 2.8% respectively. The proposed model achieves real-time processing at 30 frames per second on standard Dataset. This research contributes to intelligent transportation systems by providing a computationally efficient solution that accurately captures the temporal evolution of drowsiness symptoms, offering significant potential for reducing drowsy-related accidents through early warning systems.

The paper proposed the different scenarios that include multi approach like day time, nigh time, driver with bare eyes, with spectacles and with sunglass.