Real-time driver drowsiness detection using transformer architectures: a novel deep learning approach.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Driver drowsiness is a major cause of road accidents, leading to significant societal and economic losses.
• Existing drowsiness detection methods often lack accuracy and robustness in real-world conditions.

Purpose of the Study:

• To develop a novel and robust deep learning framework for real-time driver drowsiness detection.
• To leverage state-of-the-art transformer architectures and transfer learning for improved accuracy and reliability.

Main Methods:

• Utilized advanced data preprocessing: image normalization, augmentation, and Haar Cascade for region-of-interest selection.
• Evaluated Vision Transformer (ViT), Swin Transformer, and various transfer learning models (VGG19, DenseNet169, ResNet50V2, etc.) on the MRL Eye Dataset.
• Incorporated Class Activation Mapping (CAM) for model interpretability and real-time drowsiness scoring with alarms.

Main Results:

• Vision Transformer (ViT) and Swin Transformer achieved high accuracy rates of 99.15% and 99.03%, respectively.
• Transformer models outperformed other evaluated models in precision, recall, and F1-score.
• The system demonstrated robustness across diverse datasets (NTHU-DDD, CEW) and challenging conditions (lighting, glasses).

Conclusions:

• Transformer-based deep learning architectures offer a significant advancement for driver drowsiness detection.
• The proposed contactless system provides a reliable and efficient solution for enhancing road safety.
• The framework shows potential for widespread adoption in advanced driver assistance systems.