Wi-AM: Enabling Cross-Domain Gesture Recognition with Commodity Wi-Fi.

Area of Science:

• Intelligent multimedia technology
• Human-computer interaction
• Machine learning for signal processing

Background:

• Radio-frequency (RF)-based gesture recognition offers enhanced user privacy compared to computer vision systems.
• Integrating Wi-Fi sensing with deep learning presents new opportunities for intelligent multimedia applications.
• Current RF-based systems struggle with domain generalization and require extensive data for new environments.

Purpose of the Study:

• To propose Wi-AM, a novel framework for privacy-preserving gesture recognition.
• To address the limitations of existing systems in cross-domain performance and data requirements.
• To enable accurate gesture recognition in new domains with minimal labeled data.

Main Methods:

• Developed a multi-domain adversarial scheme to minimize domain-specific data distribution discrepancies and extract transferable gesture features.
• Implemented a meta-learning framework for rapid adaptation to unseen domains using few-shot learning.
• Utilized Wi-Fi sensing and deep learning for gesture data acquisition and analysis.

Main Results:

• Wi-AM achieves high accuracy in recognizing gestures in unseen domains, even with a single sample per gesture.
• Demonstrated average accuracies of 82.13% with one sample and 86.76% with three samples in real-world datasets.
• The multi-domain adversarial and meta-learning approaches effectively reduce the need for large datasets in new domains.

Conclusions:

• Wi-AM significantly improves cross-domain gesture recognition performance while preserving user privacy.
• The proposed framework offers a practical solution for widespread adoption of RF-based gesture recognition.
• Minimal data requirements make Wi-AM adaptable and efficient for diverse real-world scenarios.