LHAR: Lightweight Human Activity Recognition on Knowledge Distillation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Wearable Technology

Background:

• Sensor-based Human Activity Recognition (HAR) is crucial for daily life and virtual healthcare.
• Current smart wearable devices face challenges with low personalized recognition accuracy due to limited resources.
• Integrating and transmitting sensor data to the cloud reduces efficiency.

Purpose of the Study:

• To propose a Lightweight Human Activity Recognition (LHAR) framework balancing performance and complexity.
• To address the limitations of deep learning models on resource-constrained wearable devices.
• To enhance the accuracy and efficiency of personalized HAR.

Main Methods:

• Developed a teacher-student architecture with a lightweight student network using depthwise separable convolutions.
• Employed knowledge distillation from a complex teacher model to improve the student's generalization.
• Optimized knowledge distillation via ensemble learning for the teacher model and multi-channel data augmentation.

Main Results:

• The LHAR framework demonstrated superior efficiency and effectiveness compared to state-of-the-art models.
• Achieved improved recognition accuracy for personalized users on smart wearable devices.
• Validated through comparison evaluations, ablation studies, and hyperparameter analysis.

Conclusions:

• LHAR offers a viable solution for accurate and efficient personalized HAR on edge devices.
• The proposed framework effectively bridges the gap between performance and computational constraints in wearable HAR.
• LHAR shows significant potential for advancing virtual healthcare and metaverse applications.