Delay-Aware Cross-Modal Knowledge Distillation for Driver Vigilance Estimation: Toward Practical Edge Deployment.

Area of Science:

• Neuroscience and Cognitive Science
• Biomedical Engineering
• Machine Learning

Background:

• Vigilance estimation in driving is crucial for safety but faces challenges in balancing model accuracy with practical, real-world deployment.
• Electroencephalography (EEG) offers direct brain activity insights but is impractical for continuous driving monitoring.
• Other physiological signals (e.g., electrooculogram, electrodermal activity, photoplethysmography) are more practical but provide limited information.

Purpose of the Study:

• To develop a practical and accurate method for driver vigilance estimation using knowledge distillation.
• To bridge the gap between high-fidelity EEG data and more easily acquired physiological signals.
• To address the temporal misalignment inherent in multi-modal physiological signal data.

Main Methods:

• Proposed a delay-aware cross-modal knowledge distillation framework.
• Utilized EEG signals to train a teacher model, then distilled knowledge to student models using other physiological signals.
• Introduced a delay-aware soft alignment mechanism (DASA) to handle temporal differences and asynchronous dynamics between modalities.
• Designed an objective function incorporating cross-modal consistency, patch-level alignment, and smooth regularization.

Main Results:

• The proposed method demonstrated superior performance in vigilance estimation accuracy and temporal alignment compared to existing approaches.
• Validated effectiveness on the MMV and SEED-VIG datasets.
• Maintained real-time performance suitable for edge deployment.

Conclusions:

• The delay-aware cross-modal knowledge distillation method offers an effective solution for practical and accurate driver vigilance estimation.
• The DASA mechanism successfully addresses temporal misalignment in multi-modal physiological signals.
• This approach enhances the feasibility of using readily available physiological sensors for real-time driver safety monitoring.