PaperlayerView exact revision on arXiv

Source-linked AI summary

Breathfinding: A Wireless Network that Monitors and Locates Breathing in a Home

Neal Patwari, Lara Brewer, Quinn Tate, Ossi Kaltiokallio, Maurizio Bocca

arXiv:1302.3820v1cs.HC

TL;DR

Body-contact sensors and specialized radar are existing approaches to breathing monitoring, while RSS-based methods face motion interference and localization has required movement or empty-area calibration. This paper uses RSS from multiple links to estimate breathing despite movement and locate a breathing person; its tests report lower rate-estimation error and about 2 m average localization error.

  • Problem

    RSS links can be sensitive to breathing, but other movement causes larger signal changes, and prior RSS localization generally requires movement or empty-area calibration.

    Source: §I · p. 1

  • Method

    The method detects sudden RSS changes and subtracts their effects for rate estimation, while using link power spectral density at the breathing rate to estimate location.

    Source: §I · p. 1, B. Breakpoint Method · p. 4, §VI · p. 10

  • Results

    1.0 bpm average breathing-rate error, a 41% reduction from the basic method, with four of five tests above 89% acceptable rates; localization averaged about 2 m error.

    Source: §I · p. 2, C. Apartment Experiment Results · p. 5

  • Takeaways & Limitations

    RSS monitoring can estimate breathing and locate a breathing person without empty-room calibration, including when the person is otherwise stationary.

    Source: §I · pp. 1–2, §VI · p. 10

    The authors did not test localization through walls or in rubble.

    Source: §I · p. 1

Abstract

from arXiv · show

This paper explores using RSS measurements on many links in a wireless network to estimate the breathing rate of a person, and the location where the breathing is occurring, in a home, while the person is sitting, laying down, standing, or sleeping. The main challenge in breathing rate estimation is that "motion interference", i.e., movements other than a person's breathing, generally cause larger changes in RSS than inhalation and exhalation. We develop a method to estimate breathing rate despite motion interference, and demonstrate its performance during multiple short (3-7 minute) tests and during a longer 66 minute test. Further, for the same experiments, we show the location of the breathing person can be estimated, to within about 2 m average error in a 56 square meter apartment. Being able to locate a breathing person who is not otherwise moving, without calibration, is important for applications in search and rescue, health care, and security.

I. Introduction

The paper uses multi-link RSS to estimate breathing and locate a breathing person, addressing motion interference and prior requirements for movement or empty-area calibration.

I. Introduction

  • Prior RSS localization methods generally require a person to move or an empty-area calibration, leaving motionless people difficult to locate in rescue or fire scenarios.The paper notes that its method addresses this gap but is not tested through walls or in rubble.Source: §I · p. 1
  • Potential applications include baby monitoring, in-home sleep-apnea diagnosis, and elder-care context, although the authors have not developed algorithms to distinguish two people breathing in one room.The paper also presents standard wireless networks as a low-cost alternative to specialized radar devices.Source: §I · pp. 1–2
  • With 30 seconds of data, the method’s average breathing-rate error is 1.0 bpm versus 1.7 bpm for, and it estimates rates during windows containing motion.The paper attributes the improvement to identifying sudden movement and subtracting its effects.Source: §I · pp. 1–2
  • The paper demonstrates breathing monitoring with wireless devices meters, rather than centimeters, from the person and examines the improvement from multiple frequency channels over single-channel RSS.Its earlier study placed transceivers within centimeters of the subject.Source: §I · p. 2
  • Without empty-room calibration or training, the method maps breathing intensity and locates the breathing person, with about 2 m average error.The paper says it does not test localization through walls or in rubble.Source: §I · pp. 1–2

II. Experiments

  • Section II describes the testbed network, residential environment, and testing procedure.Source: §II · p. 2

A. Network

The testbed treats each transmitter–receiver–channel combination as a logical link and samples its RSS regularly. The apartment experiment uses a multichannel network in a furnished residence and tests several stationary breathing positions.

A. Network

  • Each transmitter/receiver/channel combination is a distinct logical link; full connectivity is used in the experiments but is not required for breathing monitoring.Source: A. Network · p. 2
  • RSS is sampled at regular intervals; the paper notes that sampling above 1.24 Hz is needed for Nyquist-rate newborn monitoring and leaves irregular sampling to future work.RSS values are denoted r_l[n] in dB.Source: A. Network · p. 2
  • The network uses TI CC2531 802.15.4 nodes and a multichannel TDMA protocol that cycles channels synchronously and sends received-link RSS to a sink for post-processing.Source: A. Network · p. 2

B. Apartment Experiment

  • The apartment experiment deploys 33 nodes at 0.9 m, uses four channels, and measures each link every 428 ms, or 2.3 Hz.The channels are selected to overlap least with Wi-Fi.Source: B. Apartment Experiment · p. 2
  • The testbed occupies a furnished 7 × 8 m apartment in a six-story building; the floor-plan figure marks node locations and dimensions in centimeters.Source: B. Apartment Experiment · pp. 2–3, Figure 1 · p. 3
  • Five tests lasted 3–7 minutes each, with the experimenter maintaining a metronome-set 10 bpm breathing rate at five locations—standing in the kitchen; sitting by the dining table, sofa, or toilet; or lying in bed—and moving about twice per minute.Source: B. Apartment Experiment · p. 3

C. Nap Experiment

The nap experiment tests breathing monitoring during a 66-minute bedroom nap, alongside the basic RSS-based rate-estimation approach. The basic method removes each link’s mean RSS before estimating breathing rate from the combined link spectra.

C. Nap Experiment

  • A router on the floor below and a transmitting laptop provide WiFi interference during the bedroom test.The laptop is at (2m, 0.5m).Source: C. Nap Experiment · p. 3
  • The nap experiment is a single 66-minute test in which the experimenter gets into bed and takes a nap.No separate breathing sensor provides ground truth, and the experimenter cannot maintain a constant breathing rate while sleeping.Source: C. Nap Experiment · pp. 3–4

A. Basic Method

  • The basic estimator sums power spectral densities across links and selects the frequency at the maximum of the sum as the breathing rate.The method uses the most recent N samples from each link.Source: A. Basic Method · p. 4
  • The basic method forms each link’s RSS signal by subtracting its average RSS, which the paper says removes a strong DC component.The paper reports using a window average for this mean removal when comparing against the earlier method.Source: A. Basic Method · p. 4

B. Breakpoint Method

The breakpoint method addresses motion interference by detecting sudden RSS changes and recalculating link averages around them. It applies mean removal to the resulting piecewise RSS signal before estimating breathing rate.

B. Breakpoint Method

  • Motion interference can hide breathing because abrupt movements create RSS changes that mean subtraction over a window may not remove.Breathing causes slow, periodic changes, while motion interference typically causes sudden changes and may shift a link’s average RSS.Source: B. Breakpoint Method · p. 4
  • The method detects time indices of sudden RSS changes as breakpoints, allowing link averages to be recalculated after each change.It uses a T-test comparing equal-duration groups of RSS samples before and after the time-under-test.Source: B. Breakpoint Method · p. 4
  • A breakpoint mean is piecewise constant, averaging RSS samples between the nearest breakpoints around each time index.The resulting mean-removed RSS signal is then used for breathing-rate estimation.Source: B. Breakpoint Method · p. 4
  • Fig. 4 plots mean-removed RSS for four links during the sofa experiment, with the basic method above the breakpoint method and estimated breakpoints shown.The caption identifies the comparison and plotted signals but does not report a quantitative outcome.Source: Figure 4 · p. 5

C. Apartment Experiment Results

The breakpoint method removes RSS mean drift that can obscure breathing, improving rate estimates over the basic method. Its performance depends on the breakpoint threshold, with the bed test showing a trade-off.

C. Apartment Experiment Results

  • The breakpoint method detected RSS shifts near n = 175 and 213, removing local mean drift; its PSD peaked at 0.166 Hz, near the true 0.167 Hz, while the basic method peaked at 0.10 Hz.Figure 5 compares normalized average PSD and estimated rates from the two methods with the true breathing rate.Source: C. Apartment Experiment Results · p. 5, Figure 5 · p. 5
  • 81% of breakpoint-method estimates were acceptable versus 61% for the basic method, and average breathing-rate error fell to 1.0 bpm, a 41% reduction.Four of five tests exceeded 89% acceptable estimates with the breakpoint method, while none did with the basic method.Source: C. Apartment Experiment Results · p. 5
  • With γ = 0.5, the bed test would have had 73% acceptable estimates instead of 33%, but the table test would have fared worse.The authors identify adaptive breakpoint selection as future work.Source: C. Apartment Experiment Results · p. 5

D. Nap Experiment Results

In the nap experiment, the breakpoint method produced few very low rate estimates outside detected motion periods. The authors relate estimated rate changes to sleep and waking patterns, while noting that true breathing rate was unavailable.

D. Nap Experiment Results

  • The nap experiment has no ground truth, and undetected motion can make the estimated rate rail to fmin, set to 6 bpm in these experiments.Small-amplitude step changes may escape breakpoint detection; the basic method does not remove step functions.Source: D. Nap Experiment Results · pp. 5–6
  • The nap estimates fell by approximately 20% during the first 20 minutes, remained relatively stable during sleep, rose after detected movement around minute 38, and increased as the person awoke.The authors describe these trends as consistent with published observations of sleep, movement, and waking.Source: D. Nap Experiment Results · p. 6
  • Only 9 of 793 window estimates (1.1%) were very low during periods without detected motion using the breakpoint method.There were 14 estimates below 9 bpm, five during identified motion periods.Source: D. Nap Experiment Results · p. 6
  • The basic method produced 48 estimates below 9 bpm, and its median also fell below 9 bpm during periods believed to contain movement; the breakpoint-method median never did.Figure 6 shows breakpoint-method estimates, their 90-second median, and detected movement periods; Figure 7 shows the basic-method estimates and median.Source: D. Nap Experiment Results · p. 6, Figure 6 · p. 6, Figure 7 · p. 6

IV. Breathing Localization

The paper adapts variance-based radio tomographic imaging to map breathing from wireless-link measurements. It reports that an approximate location of a stationary breathing person can be determined experimentally.

IV. Breathing Localization

  • The method computes a map of where breathing at the estimated rate is occurring and estimates the experimenter’s coordinate within the apartment.Table III is titled “Breathing mapping and localization parameters.”Source: §IV · p. 6, Table III · p. 7
  • Variance-based radio tomographic imaging identifies moving people, but a stationary breathing person does not appear in its image because noise outweighs breathing-related changes.The authors adapt the approach from for breathing mapping and localization.Source: §IV · p. 6
  • Experimental data show that an approximate location of the breathing person can be determined.Source: §IV · p. 7

A. Method

The localization method maps breathing energy from per-link measurements into an image over space. It models link values as a noisy linear combination of image values and computes a regularized estimate.

A. Method

  • The method uses each link’s PSD at the estimated breathing rate as an input to the localization method.Source: A. Method · p. 7
  • The vector of link values v is modeled as a linear combination of the spatial breathing-amplitude image x through weighting matrix W, plus noise η.Each x_k represents breathing energy from pixel k.Source: A. Method · p. 7
  • Each W weight depends on the transmitter, receiver, and pixel geometry; P_l normalizes the weights so each link’s total weight is 1.The supplied passage describes weights for pixels within an ellipse around the link.Source: A. Method · p. 7
  • Because solving for x is ill-posed, the method uses a regularized least-squares solution based on an assumed image covariance matrix G.The covariance entries use variance σ_x^2 and correlation distance δ.Source: A. Method · p. 7
  • The location estimate is the coordinate of the maximum-valued pixel in the image; localization of multiple people is left for future work.After Π is computed once, real-time image computation requires one matrix multiply with O(LP) multiplies and adds.Source: A. Method · p. 7

B. Apartment Experiment Results

Apartment tests evaluate breathing localization, while the nap experiment examines location estimates across 792 windows. The apartment method averages 2.1 m localization error, and nap estimates have 1.16 m RMS error in a smaller area.

B. Apartment Experiment Results

  • 2.1 m average localization error with basic mean removal; the breakpoint method averages 2.4 m error in the apartment tests.The authors describe the 2.1 m estimate as coarse in a 7 by 8 meter area, but typically sufficient to identify the room.Source: B. Apartment Experiment Results · p. 7
  • Hallway disturbances sometimes shifted the breathing image, but event times were not recorded, leaving the cause unverified.The authors suggest controlled future experiments to investigate these shifts.Source: B. Apartment Experiment Results · p. 7

C. Nap Experiment Results

  • 1.16 m RMS localization error across nap estimates, with their average about 1.0 m from the true chest position.The 792 location estimates cover an area about one-third the apartment's size; Figure 9 shows the estimates and their average.Source: C. Nap Experiment Results · p. 8, Figure 9 · p. 8

V. Performance vs. Size

The section evaluates breathing-rate estimation with fewer nodes and channels, using short-term variation as a proxy for accuracy. Results indicate that carefully selected node subsets can perform well, while performance generally improves with more nodes and channels.

A. Number of Nodes

  • Without ground-truth nap breathing rates, estimation performance is assessed by RMS difference from the median rate in a centered 90-second window.The paper calls this measure RMS-median and assumes short-term variation characterizes a less-accurate estimator.Source: A. Number of Nodes · p. 8
  • Average RMS-median generally decreased as node count increased, but every tested subset size had a trial below 1.0 and the lowest possible values had a shallow minimum at six nodes.The study ran 30 random trials for each subset size below 12; Figure 10 plots RMS-median against node count.Source: A. Number of Nodes · pp. 8–9, Figure 10 · p. 9
  • RMS-median was 0.72 bpm with nodes {0, 2, 4, 6}, versus 1.96 bpm with floor nodes {1, 3, 5, 7} and 3.57 bpm with {0, 1, 2, 3}.The authors suggest floor links may interact weakly with the person, while the closest nodes have no link crossing the chest.Source: A. Number of Nodes · p. 9, Table V · p. 9

B. Number of Channels

The channel experiment tests nap data using channel subsets and compares node configurations. Its results show improvement with more channels on average, while the conclusion highlights mean removal for small movements and link-PSD-based localization.

B. Number of Channels

  • With all 12 nodes, average RMS-median fell from 1.56 bpm at C = 1 to 0.91 bpm at C = 5, a 41% reduction; the four-node set achieved a 73% reduction using five channels.For nodes {0, 2, 4, 6}, C = 1 averaged 2.54 bpm and all five channels yielded 0.83 bpm; Figure 11 compares channel counts for both node sets.Source: B. Number of Channels · p. 9, Figure 11 · p. 9

VI. Conclusion

  • The conclusion identifies mean removal as a way to handle small movements and link PSD at the breathing rate as a basis for locating the breathing person.It presents these advances as supporting breathing monitoring and localization in an apartment testbed.Source: §VI · p. 10
  • Future questions include adaptive sampling to reduce traffic and energy, monitoring multiple people, and preventing surreptitious breathing surveillance over WLAN.The conclusion also raises statistical and radio-propagation models and use of other wireless hardware.Source: §VI · p. 10
Loading 1302.3820v1…