RISAR: Reconfigurable Intelligent Surfaces-Assisted Human Activity Recognition With Commercial Wi-Fi Devices

Junshuo Liu, Tiebin Mi, Xin Shi, Yunlong Huang, Zhe Li, Wei Yang, Rujing Xiong, Robert C. Qiu

TL;DR

This paper introduces RISAR, a novel system that uses Reconfigurable Intelligent Surfaces (RIS) to boost the performance of human activity recognition (HAR) using everyday Wi-Fi devices, especially in challenging low-signal environments. By leveraging RIS for enhanced spatial diversity and a new high-dimensional factor model for feature extraction, RISAR achieves state-of-the-art accuracy, demonstrating a significant leap in practical Wi-Fi-based sensing.

What Problem Does This Paper Address?

Human Activity Recognition (HAR) is crucial for applications like smart homes, healthcare, and security. Traditional Wi-Fi-based HAR systems, while promising, often struggle with limited spatial diversity from a small number of antennas, leading to poor performance in noisy or multipath-rich indoor environments. Existing methods for noise reduction and feature extraction, like Principal Component Analysis (PCA), are often insufficient for the complex, non-linear patterns found in Wi-Fi Channel State Information (CSI). This gap limits the reliability and accuracy of HAR systems in real-world scenarios.

Methods

The core of RISAR lies in its innovative combination of hardware and software:

  • Reconfigurable Intelligent Surfaces (RIS): RIS acts as a smart reflector, intelligently redirecting Wi-Fi signals to enhance spatial diversity. This is particularly beneficial in low Signal-to-Noise Ratio (SNR) conditions, effectively creating more robust communication links. The paper highlights this as the first attempt to use RIS for HAR with commercial Wi-Fi devices.
  • High-Dimensional Factor Model (HDFM): For feature extraction and noise reduction, the authors propose a novel HDFM based on random matrix theory. This method is designed to better capture the complex, non-linear temporal-spatial correlations in CSI data compared to traditional PCA.
  • Dual-Stream Spatial-Temporal Attention Network (DS-STAN): A specialized deep learning architecture is employed. It uses a dual-stream approach to process both amplitude and phase CSI features separately, then fuses them using a spatial-temporal attention mechanism. This attention mechanism mimics human cognitive processes by learning to focus on the most relevant features and time steps, significantly improving recognition accuracy.

Datasets

The experiments were conducted in three distinct environments:

  • L-shaped hallway: This environment was specifically designed to test RISAR under extremely low SNR conditions.
  • Meeting room (12x8.7 m²):
  • Office room (8x8.7 m²):

For these environments, ten typical human activities were recorded (bend, empty, fall, lie down, pick up, run, sit down, stand, stand up, walk). Participants repeated each activity 100 times, with data sampled at 1000 packets per second over 5-second intervals, creating a comprehensive dataset.

Key Insights and Ideas

  • RIS as a Game-Changer for Wi-Fi Sensing: The paper strongly demonstrates that RIS can significantly boost the performance of Wi-Fi-based sensing, particularly in challenging environments where traditional methods falter. It's not just about adding more links, but about intelligently shaping the signal environment.
  • Beyond PCA for CSI Feature Extraction: The HDFM's foundation in random matrix theory offers a more mathematically rigorous approach to denoising and feature extraction from CSI, outperforming PCA by better handling the inherent non-linearities and complex correlations.
  • Attention is All You Need (for HAR): The DS-STAN's attention mechanism proves crucial for discerning subtle activity patterns within the fused CSI data, leading to substantial accuracy gains, especially for specific activities like "stand up."
  • Amplitude and Phase Fusion Matters: Combining both amplitude and phase information from CSI signals, rather than relying on just one, leads to a significant performance boost, underscoring the richness of information available in Wi-Fi signals.

Key Results

  • Exceptional Accuracy: RISAR achieved an impressive average accuracy of 97.26% on the Office dataset and 90.83% on the L-shaped hallway dataset.
  • Significant Improvement with RIS: The DS-STAN model saw an accuracy jump from 79.17% (without RIS) to 90.83% (with RIS-optimized configuration) in the meeting room scenario.
  • HDFM Superiority: HDFM demonstrated notable improvements over PCA-based feature extraction, with accuracy gains ranging from 5.14% to 9.48% across different datasets and classifiers.
  • Attention Mechanism Boost: The attention mechanism further enhanced accuracy, pushing it to 96.39% and 97.26% for the Meeting room and Office datasets, respectively.
  • State-of-the-Art Performance: RISAR significantly outperformed existing HAR methods, achieving an accuracy of 96.39% on the Meeting room dataset, surpassing methods like LSTM, ABLSTM, and CEHAR.

Limitations and Future Work

  • Commercial Wi-Fi Device Compatibility: While the system uses commercial Wi-Fi devices, the specific implementation details (e.g., NICs, antennas) might require some adaptation for other hardware.
  • Environmental Specificity: The study was conducted in controlled experimental environments. Further validation in more diverse and dynamic real-world settings would be beneficial.
  • Future Work: The authors suggest exploring RIS technology for more complex scenarios, such as multi-person activity recognition and integration with other sensing modalities, to further revolutionize wireless sensing systems.

Summary updated

BibTeX
@article{liu2024risar,
  author = {Junshuo Liu and Tiebin Mi and Xin Shi and Yunlong Huang and Zhe Li and Wei Yang and Rujing Xiong and Robert C. Qiu},
  title = {{RISAR: Reconfigurable Intelligent Surfaces-Assisted Human Activity Recognition With Commercial Wi-Fi Devices}},
  journal = {IEEE Internet of Things Journal},
  year = {2024},
  doi = {10.1109/jiot.2024.3454223},
}
Figure