Facial Video-Based Non-Contact Stress Recognition Utilizing Multi-Task Learning With Peak Attention

Area of Science:

• Biomedical Engineering
• Computer Science
• Psychology

Background:

• Stress, anxiety, and depression are prevalent societal challenges.
• Wearable devices for stress monitoring cause discomfort due to constant skin contact.
• Accurate stress recognition is vital for effective stress management and intervention.

Purpose of the Study:

• To develop a non-contact stress recognition framework using facial videos.
• To enhance stress recognition accuracy by leveraging multi-task learning.
• To provide a comfortable alternative to wearable devices for prolonged stress monitoring.

Main Methods:

• A peak attention-based multitasking framework was developed.
• Remote photoplethysmography (rPPG) signals were extracted from RGB facial videos.
• A novel multi-task attentional convolutional neural network for stress recognition (MTASR) was employed, incorporating peak detection and heart rate (HR) estimation as auxiliary tasks.

Main Results:

• The MTASR model achieved 94.33% accuracy for stress state recognition.
• The model demonstrated 83.83% accuracy for stress level recognition.
• The proposed non-contact method outperformed existing baseline and competing approaches on the UBFC-Phys dataset.

Conclusions:

• The developed non-contact stress recognition framework offers a comfortable and effective solution.
• Multi-task learning significantly enhances feature extraction efficiency for stress recognition.
• This approach provides a promising advancement in remote physiological monitoring for mental well-being.