TCN-MAML: A TCN-Based Model with Model-Agnostic Meta-Learning for Cross-Subject Human Activity Recognition.

Area of Science:

• Pervasive computing
• Machine learning
• Signal processing

Background:

• Human activity recognition (HAR) using Wi-Fi sensing offers non-intrusive monitoring.
• Existing methods struggle with cross-subject variability and limited labeled data.
• Wearable sensors require user compliance, limiting widespread adoption.

Purpose of the Study:

• To develop a novel framework for efficient cross-subject adaptation in Wi-Fi-based HAR.
• To address data scarcity and individual differences in Wi-Fi sensing.
• To improve the generalization and accuracy of HAR systems.

Main Methods:

• Integration of Temporal Convolutional Networks (TCN) with Model-Agnostic Meta-Learning (MAML).
• Utilizing Wi-Fi Channel State Information (CSI) for device-free human motion detection.
• Evaluation using a public Wi-Fi CSI dataset with a strict cross-subject protocol.

Main Results:

• Achieved 99.6% accuracy in cross-subject human activity recognition.
• Demonstrated superior generalization and efficiency compared to baseline methods.
• Validated the framework's effectiveness in data-scarce conditions.

Conclusions:

• TCN-MAML effectively overcomes cross-subject variability and data limitations in Wi-Fi HAR.
• The framework shows significant promise for low-power, real-time HAR in IoT sensor networks.
• This approach enhances the feasibility of ambient healthcare, security, and elderly care applications.