Non-Contact Smart Sensing of Physical Activities during Quarantine Period Using SDR Technology

Muhammad Bilal Khan, Ali Mustafa, Mubashir Rehman, Najah Abed AbuAli, Chang Yuan, Xiaodong Yang, Fiaz Hussain Shah, Qammer H. Abbasi2022 · DOI

TL;DR

This paper introduces a non-contact system using Software-Defined Radio (SDR) to monitor physical activities, including falls, during quarantine periods. By analyzing wireless channel state information (WCSI), the system achieves high accuracy in distinguishing falls from other activities, offering a privacy-preserving alternative to wearables.

What Problem Does This Paper Address?

The COVID-19 pandemic necessitated widespread quarantines, leading to reduced physical activity and increased risks of non-communicable diseases and injuries from falls, especially for individuals living alone. While wearable sensors can monitor activity, they pose hygiene risks and can be uncomfortable. Existing non-contact methods like Wi-Fi and radar sensing have limitations in terms of cost, flexibility, or deployment constraints. This research aims to bridge this gap by developing a flexible, portable, and cost-effective non-contact sensing system using SDR technology for continuous physical activity monitoring, particularly for fall detection during quarantine.

Methods

The core of the system is a non-contact sensing platform that leverages Software-Defined Radio (SDR) to transmit and receive radio signals. The key innovation lies in analyzing the Wireless Channel State Information (WCSI) derived from these signals.

  • SDR Platform: The system utilizes Universal Software Radio Peripheral (USRP) devices for transmitting and receiving radio signals. These signals are modulated using Orthogonal Frequency Division Multiplexing (OFDM).
  • WCSI Extraction: The system captures the reflected radio signals from a human subject within a room environment. This multipath signal contains information about the subject's movements. The amplitude-based frequency response of the WCSI is extracted for analysis.
  • Data Preprocessing: Raw WCSI data undergoes cleaning, smoothing (using low-pass filtering), and grouping to identify correlations.
  • Feature Extraction: A statistical approach is used to extract features from the preprocessed WCSI data. These features include statistical characteristics like mean, standard deviation, variance, RMS, kurtosis, skewness, and frequency domain features like FFT, signal energy, and spectrum entropy.
  • Machine Learning Classification: Three machine learning algorithms (Fine KNN, Linear SVM, and Fine Tree) are employed to classify different physical activities: standing, walking, running, bending, and falling. The Fine Tree algorithm was chosen for its superior performance.

Datasets

The study utilized a custom dataset collected from five healthy subjects performing five different activities (standing, walking, running, bending, and falling) ten times each. This resulted in a total of 250 experiments, with each experiment lasting 10 seconds. The data was partitioned into 90% for training and 10% for testing using a 10-fold cross-validation method. No publicly available datasets were used for this specific study.

Key Insights and Ideas

  • WCSI as a Rich Data Source: The study demonstrates that subtle changes in WCSI, even in the presence of walls, can effectively capture distinct human physical activities. This highlights the potential of radio signals as a rich source of information for human activity recognition.
  • SDR's Flexibility: The use of SDR technology offers significant advantages in terms of flexibility and cost-effectiveness. Its software-defined nature allows for easy redefinition of parameters and adaptation to different wireless standards, unlike fixed hardware solutions.
  • Non-Contact Fall Detection: The system's ability to accurately detect falls without any physical contact is crucial for elderly individuals or those in quarantine, where immediate assistance might be needed.
  • Privacy Preservation: Unlike camera-based systems, this SDR approach offers a higher degree of privacy as it does not capture visual information.

Key Results

The research achieved a high accuracy of 99.7% in classifying fall activities separately from standing, walking, running, and bending using the Fine Tree machine learning algorithm. The system also demonstrated a prediction speed of approximately 72,000 observations per second and a training time of about 9 seconds.

Limitations and Future Work

  • Limited Subject Diversity: The experiments were conducted with only five healthy subjects. Future work should involve a larger and more diverse group of participants, including individuals of different ages and physical conditions, to ensure generalizability.
  • Controlled Environment: The experiments were performed in a controlled room environment. Testing in more complex and dynamic environments with multiple people and varying clutter would be beneficial.
  • COVID-19 Symptom Detection: The authors suggest that this SDR technology platform could be further investigated to detect COVID-19 symptoms like shortness of breath, coughing, and cardiac arrest.
  • Multi-Subject Recognition: Future research could focus on recognizing the activities of multiple subjects simultaneously by increasing system gain, sampling rate, and employing MIMO antennas.
  • Phase Response Analysis: Exploring the phase response of the wireless channel could offer another avenue for recognizing multi-user activities.

Summary updated

BibTeX
@misc{khan2022non,
  author = {Muhammad Bilal Khan and Ali Mustafa and Mubashir Rehman and Najah Abed AbuAli and Chang Yuan and Xiaodong Yang and Fiaz Hussain Shah and Qammer H. Abbasi},
  title = {{Non-Contact Smart Sensing of Physical Activities during Quarantine Period Using SDR Technology}},
  year = {2022},
  doi = {10.3390/s22041348},
}
Figure