Human Occupancy Detection via Passive Cognitive Radio

Jenny Liu, Huaizheng Mu, Asad Vakil, Robert Ewing, Xiaoping Shen, Erik Blasch, Jia Li

TL;DR

This paper introduces CRhodora, a novel system for human occupancy detection (HOD) using passive cognitive radio (CR) signals. By adaptively sensing and selecting the most informative RF bands, CRhodora achieves robust and efficient HOD in various environments without invasive sensors, outperforming previous deep learning approaches.

What Problem Does This Paper Address?

Traditional human occupancy detection (HOD) methods often rely on active sensing (e.g., lidar, radar) or invasive sensors (e.g., cameras), which can be expensive, power-hungry, or raise privacy concerns. While passive sensing offers a promising alternative, detecting human presence through ambient RF signals is challenging due to spectrum variations and the subtle signatures humans leave. Existing passive RF HOD systems often require scanning a wide spectrum, leading to inefficiency and high computational costs. This paper addresses the need for a passive, low-cost, privacy-preserving, and efficient HOD system that can adapt to diverse environments and spectrum fluctuations.

Methods

The CRhodora system employs a multi-stage approach:

  • RF Signal Acquisition: A reconfigurable Software-Defined Radio (SDR) is used to collect passive RF signals from enclosed spaces (study room, bedroom, car). To minimize interference, only the laptop and SDR were powered on, and cell phone signals were excluded from training and detection.
  • RF Signal Pre-Processing: The average power spectrum per frequency band is calculated.
  • Adaptive Spectrum Sensing (Feature Selection): This is a core innovation. Instead of scanning the entire spectrum, CRhodora dynamically selects the most sensitive frequency bands for HOD. Two algorithms are used:
  • PCA-based (Unsupervised): Principal Component Analysis is used to reduce dimensionality and extract principal components, with frequency bands selected based on their loadings.
  • RFE-LR-based (Supervised): Recursive Feature Elimination with Logistic Regression is employed to iteratively remove less important features and identify the most discriminative frequency bands.
  • Classifier Training: Four traditional machine learning classifiers are trained on the selected frequency bands: Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Decision Tree (DT), and Linear SVM with Stochastic Gradient Descent (SGD).
  • Online Training: The system is designed for online adaptation, where the ML model is periodically retrained with newly collected data to maintain accuracy in changing environments.

Datasets

The paper describes data collection in three distinct locations: a study room in a single-family house, a bedroom in an apartment, and a car. For each location, RF raw data was collected with and without a human subject present. A total of 1447 frequency bands were scanned. The dataset consists of full band samples, with specific numbers used for training and testing across different experiments (e.g., 300 samples for study room/bedroom, 450 for car, with varying numbers of occupied/unoccupied samples). The paper does not introduce a new benchmark dataset but uses collected data for evaluation.

Key Insights and Ideas

  • Passive CR is Viable for HOD: The research demonstrates that ambient RF signals, when intelligently processed, can effectively detect human presence without active transmission.
  • Adaptive Spectrum Sensing is Crucial: Dynamically selecting sensitive frequency bands significantly improves efficiency and robustness compared to scanning the entire spectrum. This is a key enabler for practical, low-power HOD.
  • Traditional ML Outperforms CNNs (in this context): For this specific HOD task, traditional ML classifiers trained on carefully selected features achieved comparable or better results than a CNN (from prior work) with significantly less data and computational resources.
  • Feature Selection Matters: Both PCA and RFE-LR successfully identified sensitive frequency bands, with RFE-LR generally showing superior performance, highlighting the benefit of supervised feature selection.
  • Environmental Signatures Exist: The study observed noticeable differences in power spectrums between occupied and unoccupied scenarios, with variations dependent on the environment (e.g., larger in a car than a study room).
  • Clustering of Sensitive Bands: A phenomenon of "cluster effect" was observed, where selected frequency bands tend to group around certain frequencies (e.g., 600 MHz), suggesting potential for further optimization and understanding of human RF signatures.

Key Results

  • High Accuracy: The CRhodora system achieved high detection accuracy across various scenarios, with some configurations reaching near-perfect performance (e.g., SGD with RFE-LR achieving 100% accuracy in several scenarios).
  • Efficiency Gains: By selecting a small subset of frequency bands (e.g., 10-40 bands out of 1447), the system achieved significant savings in energy and time (nearly 97.2% improvement).
  • Robustness: The system demonstrated robust performance in different locations and environments, adapting to spectrum variations.
  • RFE-LR Superiority: The supervised RFE-LR feature selection algorithm generally led to better detection performance for all tested classifiers compared to the unsupervised PCA approach.
  • SGD Classifier Performance: The SGD classifier, particularly when combined with RFE-LR, consistently showed strong performance across different scenarios.
  • Storage and Processing: The system's storage and processing requirements were evaluated and found to be suitable for implementation on well-designed embedded systems.

Limitations and Future Work

  • Location and Number Estimation: The current system detects human occupancy but does not estimate the subject's location or the exact number of individuals.
  • Environmental Variation: While adaptive, the degree of spectrum variation is location-dependent and warrants further investigation.
  • Future Work:
    • Synthesizing Human RF Signatures: Investigating the use of Generative Adversarial Networks (GANs) to synthesize human RF signatures for improved security and smart building applications.
  • More Thorough Study of Clustering: Further exploring the observed clustering of sensitive frequency bands.
  • Optimizing Parameters: The paper suggests that performance can be further improved by smartly choosing the band selection algorithm, classifier, number of bands, and training samples.

Summary updated

Abstract

Human occupancy detection (HOD) in an enclosed space, such as indoors or inside of a vehicle, via passive cognitive radio (CR) is a new and challenging research area. Part of the difficulty arises from the fact that a human subject cannot easily be detected due to spectrum variation. In this paper, we present an advanced HOD system that dynamically reconfigures a CR to collect passive radio frequency (RF) signals at different places of interest. Principal component analysis (PCA) and recursive feature elimination with logistic regression (RFE-LR) algorithms are applied to find the frequency bands sensitive to human occupancy when the baseline spectrum changes with locations. With the dynamically collected passive RF signals, four machine learning (ML) classifiers are applied to detect human occupancy, including support vector machine (SVM), k-nearest neighbors (KNN), decision tree (DT), and linear SVM with stochastic gradient descent (SGD) training. The experimental results show that the proposed system can accurately detect human subjects—not only in residential rooms—but also in commercial vehicles, demonstrating that passive CR is a viable technique for HOD. More specifically, the RFE-LR with SGD achieves the best results with a limited number of frequency bands. The proposed adaptive spectrum sensing method has not only enabled robust detection performance in various environments, but also improved the efficiency of the CR system in terms of speed and power consumption.
BibTeX
@inproceedings{liu2020human,
  author = {Jenny Liu and Huaizheng Mu and Asad Vakil and Robert Ewing and Xiaoping Shen and Erik Blasch and Jia Li},
  title = {{Human Occupancy Detection via Passive Cognitive Radio}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=WYXhsJ7iGv},
}

Loading related content...

Figure