Robust Drowsiness Detection Using Convolutional Neural Networks: A Balanced Approach to Driver Fatigue Monitoring

Drowsiness detection is a very critical concern to guarantee safety, above all in the case of driving, since it may cause serious accidents. Traditional approaches to drowsiness detection are based on physiological signals or visual cues, but they usually fail in giving accurate results or are irritatingly intrusive. In this work, we investigate the adoption of a CNN model in classifying states of drowsiness using the images taken from the publicly available repository. It consists of labelled data distinguishing between drowsy and non-drowsy states. The model was developed over a Convolutional Neural Network architecture, involving several convolutional and pooling layers with adjacent fully connected layers. The performance measurements obtained on the test set demonstrated an overall accuracy of 92.4% with a loss value of 0.18 for the proposed CNN model. The true positive rate was 95%, while the true negative rate was 98% in the confusion matrix result, which showed that this model was decently strong at differentiating drowsy states from non-drowsy ones. These results are indicative of the potential offered by CNN in real-world, real-time detection of drowsiness and can serve as an extremely accurate, nonintrusive solution to reduce risks due to drowsiness-related accidents by long chalk. The future work can be done by improving model generalization with more data involved, followed by the systems' assessment under various environmental conditions.