Abstract
Wireless sensing using Wi-Fi signals has emerged as a promising technology for non-intrusive indoor monitoring. However, existing methods often struggle with the trade-off between data granularity and computational overhead, as well as the high cost of labeled data. This paper presents a novel multi-band Wi-Fi sensing framework that bridges the "granularity gap" between fine-grained sub-7 GHz Channel State Information (CSI) and mid-grained 60 GHz millimeter-wave (mmWave) beam SNR. By employing a learnable "granularity matching" technique and an unsupervised autoencoder for pre-training, the proposed system achieves high accuracy in pose recognition and localization with minimal labeled training data.
Introduction: The Evolution of Wi-Fi Sensing
Wi-Fi sensing leverages the way wireless signals interact with the environment to detect presence, movement, and even specific gestures. Traditionally, this has relied on two data types: coarse Received Signal Strength (RSSI) or fine-grained Channel State Information (CSI). While CSI provides a detailed view of the wireless channel, it is computationally intensive and susceptible to noise. The advent of 60 GHz mmWave routers introduces a third option: "mid-grained" spatial beam SNR (Signal-to-Noise Ratio).
Despite these advancements, two major challenges persist:
- The Granularity Gap: CSI and beam SNR operate in different physical domains (frequency/time vs. spatial). Simple concatenation of these features often fails because their levels of abstraction do not align.
- The Labeling Bottleneck: Supervised learning for Wi-Fi sensing requires extensive labeled datasets, which are difficult to collect and scale across different environments.
Methods: Bridging the Gap with Granularity Matching
To address these hurdles, the researchers developed a multi-stream architecture centered on Granularity Matching (GM). Unlike standard fusion techniques that combine only the final layers of neural networks, GM utilizes a learnable fusion block. This block pairs intermediate feature maps from the CSI and beam SNR encoders at varying depths, allowing the model to establish the optimal correspondence between fine-grained and mid-grained data.
Furthermore, the framework incorporates Unsupervised Pre-training via a multi-stream autoencoder. By "listening" to unlabeled Wi-Fi traffic, the system learns to compress and reconstruct signals, capturing the fundamental characteristics of the environment without human intervention. This pre-trained model is then fine-tuned for specific tasks using a small fraction of labeled data through transfer learning.
Experimental Validation and Results
The framework was evaluated across three distinct tasks using commercial off-the-shelf (COTS) hardware: pose recognition, occupancy sensing, and indoor localization.
Key Findings
- Enhanced Performance: The GM approach significantly outperformed single-band models. In pose recognition, it achieved 94.4% accuracy, compared to 74.8% for CSI alone.
- Data Efficiency: The system maintained an 88.0% accuracy for pose recognition even when labeled data was reduced to just 10% of the total dataset.
- Cross-Band Synergy: CSI (sub-7 GHz) excels at resolving multipath interference, while beam SNR (60 GHz) provides directional spatial information. Their fusion creates a more robust environmental "image."
Conclusion
This multi-band framework demonstrates that matching feature granularity is key to unlocking the potential of heterogeneous wireless data. By reducing the reliance on labeled data, it paves the way for more scalable and easily deployable Wi-Fi sensing systems. Future research may explore cross-task transfer learning to further minimize human effort in system calibration.
This summary is based on the research paper "Multi-Band Wi-Fi Sensing with Matched Feature Granularity."
Summary updated