Wi-CHAR: A WiFi Sensing Approach with Focus on Both Scenes and Restricted Data.

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• WiFi-based human activity recognition (HAR) struggles with large data requirements and accuracy degradation in new environments.
• Existing wireless sensing methods often fail when deployed in unfamiliar domains due to domain shift.
• Data scarcity and environmental variations pose significant challenges to robust HAR systems.

Purpose of the Study:

• To introduce Wi-CHAR, a novel few-shot learning-based cross-domain activity recognition system.
• To address the limitations of current WiFi-based HAR systems, particularly their reliance on extensive data and poor generalization.
• To develop a system capable of accurate activity recognition in diverse sensing environments with limited data samples.

Main Methods:

• Wi-CHAR utilizes a dynamic sensing device selection methodology to enhance data fidelity in multi-sensor ecosystems.
• The MF-DBSCAN clustering algorithm is employed for iterative anomaly rectification and improved behavior recognition quality.
• A Re-PN module dynamically adjusts feature prototype weights for effective cross-domain sensing with limited data.

Main Results:

• Wi-CHAR achieves an average accuracy exceeding 93% in five-shot learning scenarios across various environments.
• The system demonstrates superior cross-domain recognition performance, even with minimal data in the target domain.
• Evaluations on WiAR and Widar 3.0 datasets yield high accuracy rates of 89.7% and 92.5%, respectively.

Conclusions:

• Wi-CHAR offers a robust solution for WiFi-based human activity recognition, overcoming data limitations and domain shift challenges.
• The system's adaptive mechanisms ensure high performance in specific sensing environments and with scarce data.
• Wi-CHAR achieves state-of-the-art recognition accuracy, making it suitable for real-world applications with diverse user and environmental conditions.