Introduction
Wi-Fi-based human activity recognition (HAR) represents a significant advancement for privacy-preserving smart environments, eliminating the need for intrusive cameras or wearable sensors. However, implementing HAR on commercial Wi-Fi devices, such as the Intel 5300 NIC, faces substantial hurdles. These devices typically possess few antennas, limiting spatial diversity and making signals highly susceptible to noise and multipath interference in complex indoor layouts.
The RISAR Framework
To address these limitations, the RISAR framework introduces a synergistic approach combining hardware enhancement with advanced signal processing and deep learning.
Hardware Enhancement via RIS
The system utilizes a 1-bit Reconfigurable Intelligent Surface (RIS) consisting of a 16x10 array of reflecting units. By acting as a programmable mirror, the RIS creates additional signal paths, effectively boosting spatial diversity without requiring hardware modifications to the Wi-Fi transceivers. This setup allows the system to "see" around corners or into blind spots.
Advanced Denoising and Feature Extraction
Instead of traditional linear methods like PCA, RISAR employs a High-Dimensional Factor Model (HDFM) based on the Marčenko-Pastur law. This mathematical framework treats Channel State Information (CSI) as a large random matrix, enabling the precise separation of human activity signals from background noise by analyzing the eigenvalue distribution of the covariance matrix. This approach is more principled than heuristic-based denoising.
Dual-Stream Spatial-Temporal Attention Network (DS-STAN)
The extracted features are processed by a DS-STAN architecture. This model handles amplitude and phase data through separate streams of convolutional layers and LSTMs. An integrated attention mechanism prioritizes critical time steps and subcarriers, optimizing the fusion of amplitude and phase information for final classification.
Experimental Setup and Evaluation
The framework was evaluated across three distinct indoor environments: an L-shaped hallway, a meeting room, and an office room. Data was collected for ten typical activities, including walking, falling, and sitting, with each activity repeated 100 times per participant to ensure a robust dataset.
Results and Discussion
RISAR demonstrated high precision across all test scenarios, achieving an average accuracy of 97.26% in the office environment and 96.39% in the meeting room.
A key finding was the impact of the RIS in challenging "blind spot" scenarios. In the L-shaped hallway, where signal-to-noise ratios are typically low, the inclusion of the RIS improved recognition accuracy from 79.17% to 90.83%. Furthermore, benchmarking against state-of-the-art models like CeHAR and CNN-GRU-CNN showed that RISAR consistently provides superior precision and recall by effectively leveraging both amplitude and phase components of the Wi-Fi signal.
Conclusion
RISAR successfully demonstrates that integrating RIS with commercial Wi-Fi devices can overcome traditional sensing limitations. While current results are promising, future research will focus on dynamic RIS beamforming for real-time tracking and extending the model to multi-person activity recognition in more complex architectural layouts.
Summary updated