Video-Based Stress Detection through Deep Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Stress is a significant societal health concern.
• Contact-free stress detection using video analysis offers a scalable solution.
• Existing methods may be limited by artificial factors or lack comprehensive analysis.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for non-contact stress detection.
• To investigate the combined utility of facial expressions and action motions for improved stress identification.
• To establish a benchmark dataset for stress detection research.

Main Methods:

• A two-leveled stress detection network (TSDNet) was designed.
• TSDNet processes facial expressions and action motions separately, then fuses them using an attention-based integrator.
• A custom video dataset of 2092 labeled clips was created for evaluation.

Main Results:

• TSDNet achieved 85.42% accuracy and 85.28% F1-Score, outperforming traditional feature engineering.
• Integrating both facial and action motion data improved detection performance by over 7% compared to single-modality approaches.
• The deep learning approach demonstrated feasibility and effectiveness in analyzing subtle human behaviors for stress.

Conclusions:

• The proposed TSDNet is effective for non-contact stress detection.
• Combining facial and action motion analysis significantly enhances stress detection accuracy.
• Deep learning models offer a promising avenue for objective, large-scale stress monitoring.