Comparison of Eye and Face Features on Drowsiness Analysis.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Drowsiness is a major cause of traffic accidents, leading to severe injuries, fatalities, and property damage.
• Accurate detection of drowsiness is critical for safety, especially for operators of vehicles and heavy machinery.

Purpose of the Study:

• To design and evaluate machine learning models for drowsiness recognition using human facial features.
• To analyze neural network attention mechanisms (using Grad-CAM) in interpreting drowsiness.
• To propose and assess a novel feature analysis method (KNN-Sigma).

Main Methods:

• Development of learning models for drowsiness detection using facial and eye images processed separately and in fusion.
• Implementation of Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize neural network neuron attention.
• Introduction of K-nearest neighbors Sigma (KNN-Sigma) for feature analysis.

Main Results:

• Drowsiness recognition models achieved an Area Under the Curve (AUC) of 0.814 (face), 0.897 (eye), and 0.935 (fusion).
• Processing eye images alone initially yielded better results and more reasonable Grad-CAM visualizations.
• The fusion of face and eye signals provided the highest recognition accuracy and optimal KNN-Sigma performance.

Conclusions:

• Fusion of facial and eye-based features significantly enhances drowsiness detection accuracy.
• AI models, particularly when integrating multimodal facial and eye data, show strong potential for real-time drowsiness monitoring.
• Understanding neural network attention through Grad-CAM aids in model interpretability and optimization.