AWARE-Net: A Lightweight Joint Optimization Framework for Robust Sensor-Based Human Activity Recognition

Area of Science:

• Pervasive computing
• Mobile health
• Machine learning for sensor data

Background:

• Sensor-based human activity recognition (SHAR) is crucial for pervasive computing and mobile health.
• Current deep learning methods often optimize only one aspect, hindering real-world performance.
• Challenges include balancing accuracy, noise robustness, class imbalance, and lightweight deployment.

Purpose of the Study:

• To propose AWARE-Net, a novel lightweight framework for joint optimization in SHAR.
• To address the limitations of single-dimension optimization in existing SHAR deep learning models.
• To achieve multi-objective optimization for improved real-world SHAR performance.

Main Methods:

• Utilizing TS-ResNet as a lightweight backbone encoder.
• Integrating spatiotemporal dynamic convolution for feature encoding.
• Employing a global loss function combining class-balanced loss, contrastive learning, and temporal smooth regularization.

Main Results:

• AWARE-Net demonstrates competitive performance against state-of-the-art HAR methods.
• Experiments conducted on OPPORTUNITY, PAMAP2, and USC-HAD benchmark datasets.
• The framework achieves multi-objective joint optimization effectively.

Conclusions:

• AWARE-Net offers a promising solution for efficient and robust SHAR.
• The proposed joint optimization framework overcomes previous performance bottlenecks.
• This research contributes to advancing SHAR in practical pervasive computing and mobile health scenarios.