Camera-Position-Robust Drowsiness Estimation Based on Rotated-Facial-Video Augmentation.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Biomedical Engineering

Background:

• Drowsiness detection is crucial for preventing accidents.
• Existing systems struggle with variations in camera angles.
• Eye-openness features are sensitive to camera position.

Purpose of the Study:

• To develop a camera-position-robust drowsiness estimation system.
• To improve the reliability of drowsiness detection regardless of camera placement.
• To leverage facial video augmentation for enhanced robustness.

Main Methods:

• Utilizing head-pose information for drowsiness estimation.
• Employing rotated-facial-video augmentation to simulate different camera positions.
• Integrating estimates from models trained on original and rotated data.
• Training models on single-view data and augmenting with rotated facial images.

Main Results:

• Achieved a 6.2-9.0% relative superiority in normalized mean absolute error for data captured from challenging angles (e.g., below the face).
• Demonstrated significant improvement in drowsiness estimation robustness across varied camera positions.
• The proposed method outperformed conventional methods not using rotated data.

Conclusions:

• Rotated facial video augmentation significantly enhances camera-position robustness in drowsiness estimation.
• The integrated estimation approach provides reliable drowsiness detection across diverse viewpoints.
• This system offers a promising solution for improving driver safety systems.