Two-Stream Spatial-Temporal Graph Convolutional Networks for Driver Drowsiness Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) show promise in driver drowsiness detection but struggle with real-world challenges like varying illumination, occlusions, and head poses.
• Existing methods lack the ability to differentiate subtle driver states, such as distinguishing between talking and yawning or blinking and eye closure.

Purpose of the Study:

• To introduce a novel and robust two-stream spatial-temporal graph convolutional network (2s-STGCN) for enhanced driver drowsiness detection.
• To address the limitations of current methods by incorporating spatial and temporal features from video data and improving state differentiation.

Main Methods:

• Utilized facial landmark detection to extract key facial points from real-time videos.
• Developed a two-stream spatial-temporal graph convolutional network (2s-STGCN) that processes videos as units, capturing both spatial and temporal information, as well as first and second-order dependencies.
• Employed videos, rather than consecutive frames, as the primary processing unit for driver drowsiness detection.

Main Results:

• The proposed 2s-STGCN method achieved an average accuracy of 93.4% on the YawDD dataset.
• The method demonstrated an average accuracy of 92.7% on the evaluation set of the NTHU-DDD dataset.
• Experimental results validated the feasibility and robustness of the 2s-STGCN approach in diverse conditions.

Conclusions:

• The novel 2s-STGCN offers a significant advancement in driver drowsiness detection, outperforming existing methods.
• The approach effectively handles common challenges in real-world driving scenarios, enhancing safety.
• Processing videos as units and modeling complex spatial-temporal features provides a more accurate and reliable driver monitoring system.