---
title: AirBreath Sensing Overview
url: https://www.emergentmind.com/topics/airbreath-sensing
type: topic
---

# AirBreath Sensing Overview

Searching arXiv for recent and foundational papers on AirBreath sensing and respiratory sensing modalities.
AirBreath Sensing denotes a family of respiratory sensing methods that infer breathing from perturbations in radio signals, optical and thermal measurements, geometric depth changes, magnetic coupling, mechanical strain, impedance, humidity, or exhaled gases. A foundational result showed experimentally that standard wireless networks which measure received signal strength (RSS) can be used to reliably detect human breathing and estimate the breathing rate, an application termed “BreathTaking” [1109.3898]. Subsequent work extended the field to home monitoring, smartphone thermal imaging, depth and LiDAR sensing, wearable belts and masks, optical fiber binders, breath-gas analyzers, and interaction systems. The term also acquired a distinct meaning in 6G research: “AirBreath sensing” was later introduced as a framework for protecting over-the-air distributed sensing against interference, where “breathing depth” denotes a feature-subspace dimension rather than a physiological variable [2508.11267].

## 1. Historical development and conceptual scope

The early wireless literature established the basic device-free premise: torso expansion and contraction during breathing cause small, periodic changes in the propagation environment, and these changes can be extracted from communication signals. “Monitoring Breathing via Signal Strength in Wireless Networks” demonstrated that the collective spectral content of a network of devices reliably indicates the presence and rate of breathing, even though an individual link cannot reliably detect breathing [1109.3898]. “Breathfinding: A Wireless Network that Monitors and Locates Breathing in a Home” extended the same RSS paradigm from rate estimation to calibration-free localization in a 56 square meter apartment, including sitting, laying down, standing, and sleeping scenarios [1302.3820]. “Catch a Breath: Non-invasive Respiration Rate Monitoring via Wireless Communication” then showed that respiration rate can be accurately estimated using only a single IEEE 802.15.4 compliant TX-RX pair by exploiting channel diversity, low-jitter periodic communication, oversampling, and a hidden Markov model for motion interference [1307.0084].

Later work broadened both sensing physics and deployment settings. “ThermSense: Smartphone-based Breathing Sensing Platform using Noncontact Low-Cost Thermal Camera” moved respiration sensing onto a smartphone with a low-cost thermal camera and introduced Thermal Voxel Integration and the Respiration Variability Spectrogram (RVS) [1710.05044]. “Identification of deep breath while moving forward based on multiple body regions and graph signal analysis” addressed a harder regime in which a person is walking past a global depth camera, so the breath signal is hidden within trunk displacement and deformation and the signal length is short [2010.11734]. “Remote Breathing Monitoring Using LiDAR Technology” positioned LiDAR as a remote, privacy-respecting alternative capable of monitoring inhalation/exhalation patterns, respiratory rates, breath depth, and breathlessness across various postures [2404.10970].

In parallel, the field diversified into wearables and hybrid systems. Representative examples include a zeolite-based impedimetric breath sensor based on natural clinoptilolite [1910.12104], an IoT-based respiratory motion sensor using a force sensing resistor and BLE 5 [2412.05405], a wearable singing belt with five pressure sensors [2202.01439], a contactless magnetic sensor based on magneto-LC resonance [2102.01756], a fiber-cavity CAPS chest binder [2512.03998], and a compact data-logging breath-gas analyzer that measures O\(_2\), CO\(_2\), NO, mass flow, temperature, and relative humidity [2512.02089]. This suggests that AirBreath Sensing is best understood as a transduction class rather than a single modality.

## 2. Transduction mechanisms and sensing modalities

The field can be organized by the physical quantity through which breathing is rendered observable.

| Modality | Primary observable | Representative papers |
|---|---|---|
| Wireless RSS / CSI | Multipath-induced signal fluctuation | [1109.3898], [1302.3820], [1307.0084], [2502.12114] |
| Thermal / RGB / depth / LiDAR | Temperature, displacement, point-cloud motion | [1710.05044], [2010.11734], [2210.08799], [2404.10970] |
| Reflected infrared light-wave | Intensity variation from chest motion | [2311.01367] |
| Mechanical strain / pressure / force | Chest or abdomen expansion | [2412.05405], [2202.01439] |
| Bio-impedance / magnetic / fiber optics | Electrical impedance, LC resonance, CAPS phase | [2507.04162], [2102.01756], [2512.03998] |
| Humidity / gas composition / audio | Water adsorption, exhaled gases, SCBA sound | [1910.12104], [2512.02089], [1602.07767] |

In RSS systems, the central mechanism is multipath sensitivity. When a person breathes near or in the line-of-sight of several static communicating wireless devices, the expansion and contraction of the torso causes periodic changes in the multipath environment, which affects the RSS values measured at receivers [1109.3898]. In the single-pair IEEE 802.15.4 setting, the same phenomenon is exploited through fading information, channel diversity, and oversampling [1307.0084]. In cell-free massive MIMO, uplink pilots from user equipment are repurposed so that chest motion alters the multipath propagation between the user equipment and distributed base-station antennas, with spatial diversity compensating for limited bandwidth [2502.12114].

Optical and camera-based systems use different observables. Thermal sensing around the nostrils or mouth captures temperature oscillations associated with inhalation and exhalation, while RGB, depth, and LiDAR systems recover chest or abdominal displacement. ThermSense sums temperature over a region of interest and bandpass-filters the resulting waveform [1710.05044]. The moving depth-camera system uses relative depth signals from chest, abdomen, and chest wall regions referenced to stable points such as the nose and pelvis, followed by graph signal analysis [2010.11734]. LiDAR estimates torso centroid motion from filtered point clouds and derives respiratory rate, breath depth, and breath-hold events from the resulting time series [2404.10970].

Wearable and embedded systems often transduce circumference change, force, or humidity directly. The IoT-based motion sensor uses an FSR attached to the printed circuit board so that thoracic or abdominal expansion changes force and thus resistance in a voltage divider [2412.05405]. The singing interface uses five RP thin-film bend pressure sensors placed on the lower abdomen, back waist, and ribs to quantify breath state and Breathing Dynamic Range (BDR) [2202.01439]. The clinoptilolite sensor adsorbs water vapor in exhaled breath, decreasing resistivity and impedance and producing a measurable voltage drop under a micro-watt-level square-wave AC source [1910.12104]. The DLSIB analyzer instead targets exhaled-breath chemistry and flow, integrating electrochemical, NDIR, thermal flow, and humidity sensing into a single data-logging platform [2512.02089].

## 3. Signal models, inference procedures, and feature engineering

The canonical RSS model in early wireless work is explicitly sinusoidal. In “BreathTaking,” each high-pass-filtered RSS measurement is modeled as
$$
y_l(i) = \bar{y}_l + A_l\cos(2\pi f T_s i + \phi_l) + \epsilon_l(i),
$$
with a joint maximum likelihood estimator over many links for breathing frequency, amplitudes, and phases [1109.3898]. The approximate frequency estimator searches for the strongest common spectral component across links,
$$
\hat{f} = \underset{f_{\min} \le f \le f_{\max}}{\arg\max} \sum_{l=1}^L \left| \sum_{i=1}^N y_l(i)e^{-j2\pi fT_si}\right|^2,
$$
and the presence of breathing is tested through the network-wide statistic
$$
\hat{S} \triangleq \frac{N}{L}\sum_{l=1}^L \hat{A}_l^2.
$$

Later wireless work focused on motion interference. “Breathfinding” introduced a breakpoint method that identifies sudden RSS changes using a two-sample t-test statistic and then performs piecewise mean removal before spectral estimation and localization [1302.3820]. “Catch a Breath” used mean removal, oversampling, a decimation filter with passband \(0.1\) to \(1\) Hz, and a two-state hidden Markov model with states \(S_1\) (no motion interference) and \(S_2\) (motion interference), enabling estimation only when the HMM indicates the no-motion state [1307.0084]. In cell-free massive MIMO, the pipeline combines multiple subcarriers, projects each complex-valued breathing trajectory to a single displacement dimension, and then applies Weighted Antenna Combining (WAC) using antenna weights proportional to estimated Breathing-to-Noise Ratio [2502.12114].

Camera-based and multimodal pipelines rely more heavily on region selection, denoising, and classification. The moving depth-camera method extracts six candidate respiratory channels from three ROIs and two stable points, applies outlier removal, least-squares trend removal, and bandpass filtering from \(0.167\) to \(0.667\) Hz, then uses graph signal analysis with the objective
$$
\min_{\mathbf{x}} \|\mathbf{y}-\mathbf{x}\|_2^2 + \mu \mathbf{x}^\top \mathbf{L}\mathbf{x},
$$
whose closed-form solution is
$$
\mathbf{x}^* = (\mathbf{I} + \mu \mathbf{L})^{-1}\mathbf{y},
$$
before periodicity-based channel selection and linear-kernel SVM classification [2010.11734]. ThermSense computes a one-dimensional breathing waveform from thermal voxels and then constructs the Respiration Variability Spectrogram for time-frequency analysis [1710.05044]. The thermal/RGB breathing-pattern study uses KLT feature tracking, singular value decomposition, correlation-based signal selection, moving-window PCA, continuous wavelet transform, and one-vs-one multiclass SVM classification [2210.08799].

Wearable and ancillary systems exhibit equally explicit algorithmic structure. The singing tutor averages symmetric sensors, normalizes relative to maximal exhalation, smooths each channel with a mean filter using a 150 ms window, and defines BDR as the absolute force difference between maximum and minimum during a musical phrase [2202.01439]. The SCBA audio system uses MFCC, LPC, SVD-enhanced templates, and an SVM with the cubic kernel \(k(\mathbf{u},\mathbf{y})=(0.6\cdot \mathbf{u}^\top\mathbf{y})^3\) to classify breath events and derive breathing rate and inhalation duration [1602.07767]. The bio-impedance interaction system uses a CNN1D, self-attention, LSTM, and three post-processing strategies—low-pass, front-follows-back, and majority-rule—to fire breathing-gesture events on transitions from non-null to null predictions [2507.04162].

## 4. Reported performance and experimental conditions

Reported performance varies substantially with observation time, posture, geometry, sensor placement, and the degree of motion interference.

| System | Task / condition | Reported result |
|---|---|---|
| Wireless network RSS | 20 nodes, 30 seconds data | within 0.3 breaths per minute RMS error [1109.3898] |
| Single IEEE 802.15.4 TX-RX pair | respiration rate estimation | mean error of 0.03 breaths per minute [1307.0084] |
| Wireless home network | rate + location in apartment | about 2 m average error in a 56 square meter apartment [1302.3820] |
| Depth camera while moving forward | deep-breath identification | accuracy 75.5%, precision 76.2%, recall 75.0%, F1 75.2% [2010.11734] |
| Thermal / RGB remote sensing | multiclass breathing-pattern classification | accuracy of up to 95.79% [2210.08799] |
| Reflected IR light-wave | respiratory anomaly classification | average classification accuracy of up to 96.6% [2311.01367] |
| LiDAR | five postures/orientations | respiratory rate RMSE 0.00 to 3.21 [2404.10970] |
| Cell-free massive MIMO | waveform estimation | average correlation of 0.8, compared to 0.6 for single antenna or subcarrier methods [2502.12114] |
| Fiber-cavity CAPS binder | multiple postures | RMSE of 0.91 breaths per minute; \(R^2 = 0.99\) [2512.03998] |

The wireless RSS literature is especially clear about system-design effects. In “BreathTaking,” reliable detection and frequency estimation is possible with 30 seconds of data, with RMS error in the \(0.1\)–\(0.4\) breaths/minute range, bias within \(0.1\)–\(0.2\) bpm, and fewer than \(2\%\) invalid estimates; for 50 s or more observation, the reported number of invalid estimates is zero [1109.3898]. Performance depends on node count, with at least 13 nodes needed for sub-\(0.4\) bpm RMSE, and on antenna type, with directional patch antennas providing substantially better robustness to motion outside the monitored area than omnidirectional dipoles [1109.3898]. “Catch a Breath” reported robust estimation for sampling frequencies above 3 Hz and demonstrated simultaneous monitoring of two people with a single TX-RX pair [1307.0084].

Geometry is equally important in optical and spatial sensing. LiDAR achieved its highest breathing accuracy in frontal and overhead configurations, with slightly lower performance from behind or the sides because thoracic displacement is less pronounced in those views [2404.10970]. The reflected IR system maintained robust classification over the \(0.5\) m–\(1.5\) m range, but accuracy fell at \(1.5\) m due to lower SNR and increased variance [2311.01367]. The fiber-cavity CAPS sensor was validated across sitting, supine, prone, lateral, standing, and sleep settings, with functionality reported up to 60 BrPM and validated performance across \(8\)–\(44\) BrPM [2512.03998]. This suggests that “contactless” and “wearable” should not be treated as performance guarantees independent of deployment geometry.

## 5. Wearable, embedded, and multiparametric systems

Wearable AirBreath systems emphasize low power, continuous operation, and direct coupling to respiratory mechanics. The IoT-based FSR device uses an nRF52832 MCU, a 12-bit ADC, BLE 5, a 499 k\(\Omega\) resistor in a voltage-divider circuit, a LIS2DH12 accelerometer sampled at 50 Hz, and a total device consumption of 400 \(\mu\)W; its sensing principle is
$$
V_{out}=V_{dd}\times \frac{R}{R+R_{FSR}},
$$
with \(V_{dd}=1.8\) V [2412.05405]. The paper reports low power consumption, high precision, and ease of use, and frames the accelerometer as a mechanism for disambiguating respiratory signals from body movement.

Other wearable designs target more specialized observables. The clinoptilolite sensor uses a 5.0 Vpp, 50 Hz square wave from a handheld multimeter LCD-driving signal, a true-RMS voltmeter sampled at 8–9 samples/sec, discernible voltage drops of about 100 mV per breath, and a maximum power draw of only about 14 \(\mu\)W [1910.12104]. The magnetic MLCR platform operates at 80 MHz with an optimized sensor-to-magnet distance of 5 cm and reported breathing rates of 13, 15, and 17 breaths/min for three adults aged 30, 36, and 42, while also detecting breath shortness, breath holding, and wake-versus-sleep mode transitions [2102.01756]. The fiber-cavity CAPS binder embeds SMF-28 between two fiber Bragg gratings in a sinusoidal-like pattern inside an elastic chest binder and estimates respiratory rate from CAPS phase measurements obtained with a 1550 nm laser modulated at 2 MHz and scanned at 10 Hz [2512.03998].

Breath sensing also serves as an input channel rather than only a vital sign. The singing tutoring interface uses five pressure sensors with minimum detectable force 0.196 N and response time \(<10\) ms, converts symmetric channels into lower abdomen, back waist, and rib variables, and visually maps them to a dynamic triangle during singing [2202.01439]. Among participants with musical background, the paper reports a mean pitch accuracy increase of 21.25% with \(p = 0.003\), alongside increases in lower-abdomen BDR and reductions in rib-expansion BDR for subsets of users [2202.01439]. The bio-impedance system “iBreath” uses two wet Ag/AgCl electrodes under each armpit, AD5941 at 100 kHz and 50 mV peak-to-peak, 20 Hz sampling, and F1-scores greater than 95.2% for breathing gestures, with training on five gestures requiring about 50 seconds and user-dependent calibration requiring about 2.3 minutes [2507.04162].

Multiparametric breath analysis pushes beyond mechanics to respiratory chemistry. The DLSIB platform integrates O\(_2\), CO\(_2\), NO, mass flow, temperature, and humidity sensing, with verified CO\(_2\) error \(\le 3\%\), O\(_2\) error \(\le 0.61\%\) after calibration, mass-flow accuracy \(\pm 0.1\) SLM, and temperature/RH accuracy \(\pm 0.2^\circ\)C and \(\pm 2\%\) [2512.02089]. A plausible implication is that AirBreath Sensing increasingly includes both kinematic respiration monitoring and exhaled-breath biomarker measurement.

## 6. Applications and task expansion

Respiratory rate estimation remains a core task, but the literature has expanded to deep-breath detection, breathing-pattern classification, respiratory-anomaly recognition, localization, interaction, and activity inference. The moving depth-camera system is explicitly designed to identify deep breath while the subject is walking past a global depth camera, and its 15-subject dataset uses subject-wise train/test separation repeated 1000 times [2010.11734]. The thermal/RGB remote-sensor study covers 12 predefined breathing patterns, including eupnea, bradypnea, tachypnea, hyperpnea, hypopnea, Kussmaul breathing, Cheyne-Stokes, Biot’s breathing, and apnea, with an accuracy of up to 95.79% using a one-vs-one multiclass SVM [2210.08799]. The reflected infrared light-wave system classifies eight classes—Eupnea, Apnea, Tachypnea, Bradypnea, Hyperpnea, Hypopnea, Kussmaul’s, and Faulty Data—using four handcrafted features and decision-tree or random-forest classifiers [2311.01367].

Localization and environmental integration are distinctive in wireless systems. “Breathfinding” estimates where breathing is occurring in a home, reports about 2 m average localization error in a 56 square meter apartment, and argues that locating a breathing person who is not otherwise moving is important for search and rescue, health care, and security [1302.3820]. LiDAR emphasizes privacy-respecting operation because it produces sparse 3D point clouds rather than identifying images, and it explicitly targets inhalation/exhalation patterns, respiratory rates, breath depth, and breathlessness across seated and supine scenarios [2404.10970]. ThermSense emphasizes mobile use, real-time processing on a smartphone, and downstream use of the Respiration Variability Spectrogram for machine-learning tasks such as stress detection or emotion recognition [1710.05044].

Several papers repurpose breath as a behavioral or control signal. “iBreath” defines a gesture vocabulary of single-click, double-click, triple-click, and SOS patterns, with median gesture durations from 3.5 to 5.3 seconds and user preference favoring single-click over triple-click [2507.04162]. “i-Mask: An Intelligent Mask for Breath-Driven Activity Recognition” uses in-mask temperature and humidity to classify running, walking, sitting, and sleeping, with kNN accuracy of 96.4% and lower performance for SVM at 74.4% [2509.04544]. These systems do not treat respiration merely as a passive biosignal; they treat inhalation and exhalation patterns as structured input for human-computer interaction and activity recognition.

## 7. Limitations, misconceptions, and the 6G terminological extension

A recurrent misconception is that any single sensor stream is sufficient if breathing is periodic enough. The wireless RSS literature rejects that simplification in two different ways. “BreathTaking” states that an individual link cannot reliably detect breathing and that reliable detection emerges from the collective spectral content of many links [1109.3898]. By contrast, “Catch a Breath” shows that a single IEEE 802.15.4 compliant TX-RX pair can work accurately, but only when channel diversity, low-jitter periodic communication, oversampling, decimation, and HMM-based motion gating are added [1307.0084]. These results are not contradictory; they indicate that observability depends on signal design, diversity, and interference management.

Motion interference is the dominant technical obstacle across modalities. Wireless systems describe motions other than breathing as generally causing larger changes in RSS than inhalation and exhalation, motivating breakpoint detection and HMM state estimation [1302.3820]. The depth-camera system treats gross trunk movement as the main confound and addresses it with relative-depth measurements and graph signal analysis [2010.11734]. The IoT FSR device includes an accelerometer specifically to help discriminate respiratory-induced signals from other movements [2412.05405]. The fiber-cavity CAPS sensor interprets sudden phase baseline shifts as body movement or posture changes and uses them for artifact detection [2512.03998]. A second misconception is that privacy-preserving sensing is necessarily low-information; LiDAR, thermal imaging, and reflected infrared each claim privacy advantages while still reporting useful respiratory metrics or classification accuracy [2404.10970], [1710.05044], [2311.01367].

The term “AirBreath sensing” also requires disambiguation. In “AirBreath Sensing: Protecting Over-the-Air Distributed Sensing Against Interference,” the phrase no longer denotes respiration monitoring; it denotes a spectrum-efficient framework that cascades feature compression and spread spectrum under a fixed bandwidth constraint \(S\times G \le D\) [2508.11267]. The central design variable, “breathing depth,” is defined as the feature subspace dimension \(S\), and the paper characterizes sensing accuracy through classification discriminant gain, including the surrogate
$$
\tilde{\phi}(S)=\frac{|\mathcal{K}|\psi(S)}{\tilde{\sigma}^2 S + 1}.
$$
This later usage is conceptually separate from physiological AirBreath sensing, although both involve signal extraction from distributed sensing systems. Future work in the physiological branch is repeatedly framed around broader clinical validation, robustness under uncontrolled motion, adaptive sensor placement, and long-term unobtrusive deployment [1109.3898], [1302.3820], [2412.05405], [2512.03998].

Source: https://www.emergentmind.com/topics/airbreath-sensing