Energy-aware embedded classifier design for real-time emotion analysis.

Area of Science:

• Biomedical Engineering
• Computer Science
• Affective Computing

Background:

• Real-time human emotion classification using bio-signals is crucial for various applications.
• Existing wearable devices face challenges in adapting to strict energy constraints for continuous emotion analysis.
• Developing energy-efficient embedded systems for emotion detection remains a significant hurdle.

Purpose of the Study:

• To propose an embedded classifier for real-time emotion classification on wearable devices.
• To develop an energy-budget aware feature selection algorithm for different operating modes.
• To optimize emotion classification accuracy under constrained energy budgets.

Main Methods:

• An embedded classifier system designed for variable energy-budgeted modes.
• An offline training phase incorporating an energy-budget aware feature selection algorithm.
• Evaluation of classification accuracy and power consumption across different energy modes.

Main Results:

• Classification accuracy for arousal ranged from 95% to 75%, and for valence from 89% to 70% across modes.
• The system demonstrated a trade-off between classification accuracy and power consumption.
• Significant increase in battery life, up to 7.7 times (146.1 to 1126.9 hours), was achieved.

Conclusions:

• The proposed embedded classifier effectively achieves real-time emotion classification within energy constraints.
• The energy-budget aware approach enables adaptable performance for wearable systems.
• This work significantly enhances the practical viability of bio-signal-based emotion detection in low-power devices.