Driver Drowsiness Estimation by Parallel Linked Time-Domain CNN with Novel Temporal Measures on Eye States

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Driver drowsiness is a significant factor in road accidents.
• Existing driver drowsiness estimation systems often lack accuracy or real-time capabilities.
• Vision-based systems offer a non-intrusive method for monitoring driver alertness.

Purpose of the Study:

• To develop a robust and accurate vision-based driver drowsiness estimation system.
• To propose a novel stage-by-stage approach for improved drowsiness detection.
• To introduce new temporal measures for enhanced eye state analysis.

Main Methods:

• A stage-by-stage system processing driver images frame-by-frame.
• Calculation of eye-related features and temporal measures.
• Utilizing time-domain convolution with a parallel linked structure for drowsiness level estimation.
• Introduction of Average Eye Closed Time (AECT) and Soft Percentage of Eyelid Closure (Soft PERCLOS) as novel metrics.

Main Results:

• The system achieved a high accuracy of 95.86% on a real-world driving dataset.
• A low Mean Absolute Error (MAE) of 0.4007 was recorded.
• The proposed stage-by-stage approach outperformed end-to-end systems.

Conclusions:

• The developed vision-based system effectively estimates driver drowsiness.
• The novel temporal measures (AECT and Soft PERCLOS) significantly contribute to drowsiness detection accuracy.
• This system holds potential for enhancing road safety by mitigating fatigue-related accidents.