Respiration and Activity Detection Based on Passive Radio Sensing in Home Environments

Qingchao Chen, Yang Liu, Bo Tan, Karl Woodbridge, Kevin ChettyIEEE Access 2020 · OpenReview

Abstract

The pervasive deployment of connected devices in modern society has significantly changed the nature of the wireless landscape, especially in the license-free industrial, scientific, and medical (ISM) bands. This paper introduces a deep learning-enabled passive radio sensing method that can monitor human respiration and daily activities by leveraging unplanned and ever-present wireless bursts in the ISM frequency band. By treating radio bursts as pseudo-noise waveforms, the system eliminates the requirement for signal demodulation or decoding. Instead, it leverages phase and frequency components in conjunction with cross ambiguity function (CAF) processing and a Deep Transfer Network (DTN). Experimental results using 2.4GHz WiFi signals demonstrate the capability for human respiration detection (including through-the-wall) and classifying complex human motions such as standing, sitting, and falling with high accuracy.

Introduction

As healthcare shifts toward "Aging in Place" and smart home integration, there is a growing need for continuous, non-invasive health monitoring. Current solutions often face a "Goldilocks" problem: wearable sensors are frequently forgotten or uncomfortable for the elderly, while camera-based systems raise significant privacy concerns.

Furthermore, the 2.4GHz ISM band is increasingly congested due to the explosion of IoT devices. Most existing radio-sensing methods are "active," meaning they transmit their own signals, further contributing to spectrum interference. This research proposes a passive sensing framework that listens to existing wireless traffic, turning environmental "noise" into a source of healthcare informatics. By treating ambient WiFi signals as opportunistic "passive radar," the system can track respiration and classify complex activities like falling without adding to spectrum congestion.

Methodology

The proposed system utilizes a Software Defined Radio (SDR) setup to capture 802.11 (WiFi) bursts. The core innovation lies in treating these bursts as "pseudo-noise" waveforms, comparing the direct signal from a router to the reflected signal from a human body. The processing pipeline is divided into two main branches:

Proposed methodology for passive radio sensing, showing the dual-path processing for respiration and activity detection.

Phase-Sensitive Detection for Respiration

Standard radar techniques struggle with the fine movements of a chest wall due to coarse range resolution in WiFi bandwidths (approx. 17 meters). To overcome this, the researchers leverage the Cross Ambiguity Function (CAF) to extract the phase of the reflected signal. A Hampel filter is applied to remove outliers, and background static reflections are subtracted to isolate the quasi-periodic phase shifts caused by breathing.

Deep Transfer Network for Activity Recognition

For complex motions like sitting, standing, or falling, the system generates micro-Doppler (μ-D) signatures—time-frequency representations of movement. These signatures are treated as 2D images and processed using a Deep Transfer Network (DTN) based on AlexNet. By fine-tuning a model pre-trained on ImageNet, the system can effectively learn to recognize human activities even with a relatively small set of radio signatures.

Experimental Evaluation

The system was evaluated using two primary experimental setups:

  • Respiration Detection: Data was collected in four line-of-sight (LoS) positions and two through-wall positions (behind a 33cm brick cavity wall).
  • Activity Recognition: A dataset of six daily activities was compiled: picking up an object/standing up (M1), sitting (M2), standing from a chair (M3), falling (M4), standing after a fall (M5), and getting out of bed (M6).

Key Results

  • Respiration: The system successfully extracted breathing rates in both LoS and through-wall scenarios, with frequency analysis showing a clear peak at ~0.3 Hz, consistent with normal human respiration.
  • Activity Accuracy: The fine-tuned AlexNet (DTN) achieved a peak accuracy of 98.7% when trained on 60% of the dataset.
  • Data Efficiency: Even with only 20% of the data used for training, the DTN maintained an 86.3% accuracy, significantly outperforming traditional Support Vector Machines (SVM), which achieved only 32.9% under the same conditions.

Discussion and Conclusion

This research demonstrates that ambient WiFi signals can be effectively repurposed for high-precision health monitoring. Key insights include:

  • Phase over Range: Phase variations are a much more sensitive metric for "signs-of-life" detection than traditional time-of-flight measurements in indoor environments.
  • Transfer Learning: The success of the DTN highlights that low-level features learned by computer vision models (edges, textures) are remarkably transferable to radio frequency micro-Doppler signatures.
  • The "Zero-Doppler" Signature: Falling (M4) often lacks a "zero-Doppler" line because the body shields the direct interference from the transmitter upon hitting the floor, providing a unique visual cue for the AI.

While promising, challenges remain for real-world deployment, particularly in multi-target scenarios where disentangling signals from multiple people is difficult. Future work will explore scaling data collection and optimizing performance across different wireless waveforms, such as 5G and Bluetooth.

Summary updated

Abstract

The pervasive deployment of connected devices in modern society has significantly changed the nature of the wireless landscape, especially in the license free industrial, scientific and medical (ISM) bands. This paper introduces a deep learning enabled passive radio sensing method that can monitor human respiration and daily activities through leveraging unplanned and ever-present wireless bursts in the ISM frequency band, and can be employed as an additional data input within healthcare informatics. Wireless connected biomedical sensors (Medical Things) rely on coding and modulating of the sensor data onto wireless (radio) bursts which comply with specific physical layer standards like 802.11, 802.15.1 or 802.15.4. The increasing use of these unplanned connected sensors has led to a pell-mell of radio bursts which limit the capacity and robustness of communication channels to deliver data, whilst also increasing inter-system interference. This paper presents a novel methodology to disentangle the chaotic bursts in congested radio environments in order to provide healthcare informatics. The radio bursts are treated as pseudo noise waveforms which eliminate the requirement to extract embedded information through signal demodulation or decoding. Instead, we leverage the phase and frequency components of these radio bursts in conjunction with cross ambiguity function (CAF) processing and a Deep Transfer Network (DTN). We use 2.4GHz 802.11 (WiFi) signals to demonstrate experimentally the capability of this technique for human respiration detection (including through-the-wall), and classifying everyday but complex human motions such as standing, sitting and falling.
BibTeX
@inproceedings{chen2020respiration,
  author = {Qingchao Chen and Yang Liu and Bo Tan and Karl Woodbridge and Kevin Chetty},
  title = {{Respiration and Activity Detection Based on Passive Radio Sensing in Home Environments}},
  booktitle = {IEEE Access},
  year = {2020},
  url = {https://openreview.net/forum?id=w8BaGmoiLDK},
}
Figure