Dual-Stream Contrastive Learning for Channel State Information Based Human Activity Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• WiFi-based human activity recognition (HAR) is crucial for health monitoring but faces challenges with data privacy and the limitations of existing supervised deep learning models.
• Current contrastive learning methods, primarily developed for computer vision, are not optimized for WiFi Channel State Information (CSI) data.
• The need for privacy-preserving and data-efficient HAR methods is growing.

Purpose of the Study:

• To propose a novel dual-stream contrastive learning model, DualConFi, for self-supervised learning on raw WiFi CSI data.
• To address the data-hungry nature and privacy concerns associated with traditional supervised HAR methods.
• To develop a model capable of extracting highly-discriminative spatiotemporal features from WiFi CSI data.

Main Methods:

• Developed DualConFi, a dual-stream contrastive learning framework processing raw WiFi CSI data.
• Incorporated separate channel and temporal streams to learn distinct feature representations.
• Utilized a mutual information constraint for self-supervised learning on unlabeled data.

Main Results:

• Demonstrated the effectiveness of DualConFi on three public CSI datasets across linear evaluation, semi-supervised, and transfer learning settings.
• Showcased favorable performance of DualConFi compared to challenging baseline methods.
• Verified the efficacy of learned features through analysis of various data transformation functions.

Conclusions:

• DualConFi offers a robust and effective self-supervised approach for WiFi-based HAR using raw CSI data.
• The model successfully extracts discriminative spatiotemporal features, outperforming existing methods.
• This work advances privacy-preserving HAR by leveraging unlabeled data efficiently.