The need for precise and real-time detection systems is highlighted by the fact that drowsiness while driving is a significant factor in traffic accidents. The paper evaluates how well five deep learning models such as MobileNetV2, MobileNetV3 Small, and EfficientNet versions B0, B3, and B7 for classify driver fatigue based on yawning behavior and eye condition. Models were trained and evaluated on the publicly available Yawn Eye dataset that can be found on Kaggle that consists of 7,250 labeled images in four categories, closed eye, open eye, yawn, and no yawn. Data augmentation methods of image rotation, horizontal flipping and variation of brightness were used to augment the model generalization. To make lightweight convolutional neural networks applicable to real-time, an image classification method was used, which was based on two important fatigue predictors, namely, eye closure and yawning. Accuracy of classification and F1-scores were used to measure model performance. MobileNetV3 Small recorded the highest accuracy of 94.69% followed by EfficientNet-B0 at 93.76%, EfficientNet-B7 at 93.30% and EfficientNet-B3 at 92.38%. Even though the overall accuracy of MobileNetV2 was lower, 88.22%, it showed good F1-scores of 0.98% on both open and closed-eye. These findings suggest that MobileNetV3 Small is a good compromise between the accuracy and the computational efficiency and thus it is highly applicable in driver drowsiness detection systems in real-time.
Vimala et al. (Sun,) studied this question.