Source-linked AI summary
Improving BLE-Based Passive Human Sensing with Deep Learning
Giancarlo Iannizzotto, Lucia Lo Bello, Andrea Nucita
TL;DR
WiFi-based passive human sensing faces interference, deployment, and power-related drawbacks, while BLE has sensing limitations including coarse RSSI access and unavailable channel information. The paper applies a deep convolutional neural network to RSSI data from commercial BLE devices for occupancy detection in a challenging room. The approach achieves 99.74% accuracy and a 100% f1-score, outperforming the compared state-of-the-art method.
Problem
BLE is a promising WiFi alternative for passive human sensing, but coarse RSSI sampling, hidden transmission channels, and absent native CSI constrain signal analysis.
Method
A Deep Convolutional Neural Network analyzes RSSI samples from commercial BLE 4.x beacons and receivers to extract occupancy information.
Results
99.74% accuracy and a 100% f1-score were achieved, outperforming the state-of-the-art method evaluated in the same general scenario.
Takeaways & Limitations
Commercial BLE devices and deep learning can support accurate passive occupancy detection in a fairly general, articulated environment without direct line-of-sight restrictions.
Takeaways & Limitations
The approach has intrinsically low detection speed: experiments reduced sampling to 2 Hz, which may be insufficient for rapidly traversing a small area.
Abstract
from arXiv · showhide
Passive Human Sensing (PHS) is an approach to collecting data on human presence, motion or activities that does not require the sensed human to carry devices or participate actively in the sensing process. In the literature, PHS is generally performed by exploiting the Channel State Information variations of dedicated WiFi, affected by human bodies obstructing the WiFi signal propagation path. However, the adoption of WiFi for PHS has some drawbacks, related to power consumption, large-scale deployment costs and interference with other networks in nearby areas. Bluetooth technology and, in particular, its low-energy version Bluetooth Low Energy (BLE), represents a valid candidate solution to the drawbacks of WiFi, thanks to its Adaptive Frequency Hopping (AFH) mechanism. This work proposes the application of a Deep Convolutional Neural Network (DNN) to improve the analysis and classification of the BLE signal deformations for PHS using commercial standard BLE devices. The proposed approach was applied to reliably detect the presence of human occupants in a large and articulated room with only a few transmitters and receivers and in conditions where the occupants do not directly occlude the Line of Sight between transmitters and receivers. This paper shows that the proposed approach significantly outperforms the most accurate technique found in the literature when applied to the same experimental data.
1. Introduction
Passive human sensing collects information about people without requiring them to carry devices or actively participate. This work motivates BLE as an alternative to WiFi and applies a deep neural network to detect occupants in challenging rooms without strict placement or motion constraints.
- PHS foundations: Passive Human Sensing gathers information about human presence, motion, or activities without requiring sensed people to carry devices or participate actively.It exploits deformations in radio-frequency signals caused by human bodies and their reflected or refracted components.
- Motivation for BLE: WiFi-based PHS is effective but suffers from coexistence, collision-domain, and power-consumption drawbacks.WiFi communication and sensing can interfere when performed together or in nearby networks, while WiFi is not designed for low power consumption.
- Motivation for BLE: BLE offers an alternative through Adaptive Frequency Hopping, energy efficiency, broad device integration, and simple, flexible deployment.BLE devices are small, minimally invasive, less expensive than other solutions, and typically provide approximately 20–30 m indoor range.
- BLE research gap: BLE-based PHS remains limited by low-rate RSSI access, unavailable transmission-frequency information, and the absence of native CSI measurements.These constraints make it difficult to match fast human dynamics and distinguish RSSI variability across channels.
- Proposed approach: The proposed DNN analyzes BLE signal deformations to improve PHS accuracy and flexibility in a large, articulated room.The system uses only a few standard BLE 4.x beacons and receivers, while occupants face no specific location or motion restrictions.
- Contributions: The work contributes a COTS BLE 4.x architecture, a deep convolutional network for RSSI sequences, and validation in a non-rectangular student laboratory with non-line-of-sight regions.The architecture uses stand-alone beacons and receivers built with Arduino Zero or Raspberry Pi cards equipped with BLE adapters.
2. Materials and Methods
The proposed BLE-based PHS architecture uses multipath signal deformations and a convolutional DNN to detect occupants without requiring direct transmitter–receiver occlusion. It combines median RSSI subsampling with receiver-side aggregation to improve robustness despite BLE synchronization, channel-access, and sampling constraints.
- Sensing principle: BLE PHS detects occupants from signal deformations caused when human bodies obstruct or otherwise affect transmitter–receiver propagation paths.Non-Line-of-Sight paths may involve reflection or refraction before or after traversing the body.
- Protocol constraints: BLE 4.x exposes only coarse RSSI measurements and hides the transmission channel, producing RSSI variability that requires an effective classifier.The application cannot reliably separate samples by channel because AFH-selected frequencies are unavailable.
- Network deployment: The architecture uses beacons that illuminate the monitored area and receivers positioned so each occupant affects at least one received transmission, including reflected paths.Occupants need not lie directly on the Line of Sight between a beacon and receiver.
- Signal preprocessing: Because BLE devices are unsynchronized, the system sacrifices temporal resolution by replacing k nearby RSSI messages with their median representative sample.This reduces abrupt AFH-related RSSI changes while lowering the PHS sampling frequency by a factor of k.
- Classification pipeline: Receiver-specific Median RSSI Sample rows are timestamped, merged centrally, ordered by time, and classified by a DNN for occupant detection.Each row contains contemporaneous representative samples from the different beacons.
- Classifier design: The proposed pure convolutional network uses three 1D convolutional layers, batch normalization, global average pooling, and a final dense output layer.The architecture is reported as outperforming the tested dense, LSTM, and two-layer convolutional alternatives in all tests.
3. Results
The experiments evaluated BLE-based occupancy detection using commercial devices, balanced datasets, training convergence, and comparisons with a prior method under degraded network configurations.
- Experimental setup: The testbed used seven battery-operated BLE beacons, Raspberry Pi receivers, and freely moving occupants in a laboratory environment.The room was L-shaped and included six workstations; occupants could stand and move freely.
- Data: The dataset contained 10,412 samples, split into 7,288 training-and-validation samples and 3,124 test samples with approximately balanced classes.Presence and Non-presence counts differed by less than 10% in both datasets.
- Training: Training converged to 100% accuracy and 0% error in both training and validation after an initial oscillation phase.Figure 8 tracks accuracy, validation accuracy, error, and validation error during training.
- Test results: The trained model achieved 0.9974 test accuracy, or 99.74%, on samples unknown during training and validation.The confusion matrix was used to assess discrimination on the held-out test dataset.
- Comparisons: The proposed approach was compared with a state-of-the-art BLE occupancy method and evaluated under progressively reduced beacon and receiver configurations.Table 2 reports performance after dropping receivers and beacons.
4. Discussion
The discussion positions the approach as accurate in a flexible, non-ideal setting while identifying deployment practicality and two important limitations.
- Evaluation setting: The evaluation used a concave laboratory, variable occupancy patterns, few occupants, and BLE interference rather than an idealized room configuration.Some beacon-receiver pairs lacked direct Line-of-Sight visibility.
- Performance: The approach achieved 99.74% accuracy and 100% f1-score, outperforming its direct competitor by 3% on both accuracy and f1-score.The authors report this as the best performance in the literature for the test scenario.
- Robustness: Dropping half the beacons or two of three receivers only marginally reduced detection accuracy while breaking Line-of-Sight visibility for most device pairs.This supports operation without many transmitters and receivers.
- Deployment: The detection architecture was lightweight enough to run on an embedded Raspberry Pi 3B+ computer.The authors selected the more powerful receiver platform to support future distributed processing.
- Limitations: The architecture’s sampling rate was reduced to 2 Hz, limiting detection of rapidly traversing people in small areas.The BLE 4.x advertising limit is 50 messages per second, with further division by the subsampling factor k.
- Limitations: The system requires a one-time calibration phase lasting about 3–4 hours with manual recordings of occupant presence and absence.This training requirement is shared with most approaches in the literature.
5. Conclusions
The paper presents a commercial BLE passive sensing architecture using a Deep Convolutional Neural Network to extract occupancy information from RSSI samples. In experiments, it outperformed the compared state-of-the-art method in a general scenario.
- Contribution: The architecture combines commercial BLE 4.x beacons, Raspberry Pi receivers with standard BLE modules, and a Deep Convolutional Neural Network.The network analyzes and classifies beacon RSSI samples to extract occupancy information.
- Assessment: The paper describes the architecture’s advantages and limitations and provides experimental results assessing its validity.
- Conclusion: The proposed approach achieved 99.74% accuracy and a 100% f1-score in a fairly general experimental scenario.The authors report it as the best performer with respect to the literature.