DrowsiGuard: Leveraging Deep Learning for Accurate and Efficient Drowsiness Detection

Drowsy driving remains a significant cause of traffic accidents that claims lives, cause injuries, and cost losses in terms of resources monthly, weekly, and daily. This is specifically lousy news as systematic attempts at regulating the problem and public health campaigns continue to fall short of preventing deadly accidents resulting from driver drowsiness. Previous detection techniques use bio-signals such as EEG, ECG, or vehicle parameters such as steering angle, lane drift, etc, which are intrusive, expensive, and influenced by environmental conditions; hence, they are unsuitable for real-world consumer vehicles. Real-time, noninvasive drowsiness detection is also an excellent opportunity for machine learning. However, it is difficult to achieve high accuracy, speed, and insensitivity to the changes of driving conditions using classical methods. This paper proposes DrowsiGuard, a deep learning-based system that employs the ResNet-50 model for driver drowsiness detection. We utilize the DDD dataset with engineered low-level features to compare the performance of DrowsiGuard to other models. Detailed results of the experiment depict that the proposed DrowsiGuard method is far superior to the set baselines regarding accuracy, precision, or recall values.