2017 IEEE 19th International Conference on e-Health Networking, Applications and Services (Healthcom)

TinySense: Multi-User Respiration Detection using Wi-Fi CSI

Signals
Pei Wang, Bin Guo, Tong Xin, Zhu Wang, Zhiwen Yu
Northwestern Polytechnical University, Xi’an, China
[email protected]

Abstract—Respiration rate plays an important role in propagates from the transmitter to the receiver and reflects
human health monitoring. Traditional respiration rate the combined effects of the surrounding objects (e.g.,
monitoring techniques usually require users to wear some scattering, fading, and power decay with distance). For
special equipment, which is not convenient for the elderly example, the Intel 5300 NICs report 30 subcarriers in Wi-
and the baby. Recently, Wi-Fi based respiration detection Fi signal, and the CSI of each subcarrier contains
technique has attracted much attention due to its device-free information of attenuation and phase shift in the wireless
and low-deployment-cost. However, most existing studies channel. Therefore, CSI contains rich information and is
focus on respiration detection in experimental environments, more sensitive to environmental variances caused by
without considering the impact of people around (it often
moving objects. Several notable studies on pervasive
occurs in our daily life), therefore, if there are several people
in the system, their detection will fail. To address this open
sensing have been conducted using CSI, such as high-
accurate human localization, human activity recognition,
issue, we propose TinySense, a novel approach that can
detect multiple persons’ respiration at a time. In particular,
and crowd counting [2] [3] [4]. Meanwhile, there are many
we use multiple TX-RX antenna pairs to capture the Wi-Fi works focus on human respiration monitoring using Wi-Fi
Channel State Information (CSI), filter out the data whose CSI signals. Zhang et al. [1] introduced the Fresnel Zone
time-of-arrival (TOA) is bigger than a truncation threshold Model in free space, based on which they investigated the
and remove subcarriers that are greatly affected by the impact of human respiration on the receiving RF signals
multi-path effect. As a result, we can obtain the respiration and developed the theory to relate one’s breathing depth,
data of each person from the mixed received signal. location and orientation to the detectability of respiration.
Experiments demonstrate the effectiveness of our approach However, their model requires an experiment environment
on two-user respiration detection. without persons around the user, and if there are two
subject respiration in their system they use Power Spectral
Keywords-component; Respiration Detection; Wi-Fi Density (PSD) to separate the respiration data of each
Sensing; Channel State Information; Noise Removal; person, but it is difficult to tell which breathing
I. INTRODUCTION information is belonging to whom, i.e., only one person’s
Continuous respiration monitoring plays an important respiration can be detected at a time. In other word, the
role in healthcare. An abnormality in respiration rate is a existing respiration detection systems are not reliable on
sensitive indicator of some physiologic diseases, such as multiple user respiration detection.
interstitial lung disease (too fast) or drug overdose (too In this paper, we propose TinySense, a novel approach
slow). Meanwhile, respiration monitoring is an effective for multi-user respiration detection. In particular, by
way to monitor human sleep. putting a receiver besides each user, what we need to do is
During the past few years, researchers make numerous mitigating the effect caused by other people. To address
attempts to develop methods for Smart Home construct this issue, we filter out the data whose time of arrive (TOA)
using cameras, wearable sensors or Radio Frequency (RF) is bigger than a truncation threshold and remove
devices [17] [18] [19]. Human respiration is a significant subcarriers that are greatly affected by multi-path. In this
demand of Smart Home, However, Camera-based way, our system can detect multiple persons’ respiration
approaches are sensitive to the lighting conditions and may at a time, which is more suitable in daily life. In addition,
cause privacy concerns. Wearable sensor based different from the existing system using one pair of TX-
approaches are inconvenient as users have to wear certain RX antenna, we use multiple TX-RX antenna pairs to
devices. RF-based approaches, ranging from Doppler radar, enhance the ability to capture human respiration.
Ultra-Wideband (UWB) radar to Universal Software
Radio Peripheral (USRP) based solutions, are quite
accurate but too expensive and the devices are not widely II. RELATED WORKS
deployed in our daily living environments.
Recently, Wi-Fi based pervasive sensing become A. Respiration Detection
popular due to its popularity and low-deployment-cost. During the past few years, quite a number of studies
Especially when Channel State Information (CSI, i.e., a have been devoted to respiration detection. The most
fine-grained information regarding Wi-Fi communication) direct way usually in the form of a smart sleeping mat
[1] becomes available. CSI describes how the signal based on pressure sensor arrays like [5] and [6]. Such a

978-1-5090-6704-6/17/$31.00 ©2017 IEEE

 2017 IEEE 19th International Conference on e-Health Networking, Applications and Services (Healthcom)

system can provide fine-grained information like arriving algorithm to filter the effect of person around,
respiration, body motion, and even heart rate. However, realizing a multi-user respiration detection system.
this kind of device has a high cost. For example, the price
of the Tanita [6] is over 500 USD. Some studies attempt III. PROBLEM ANALYSIS AND SYSTEM
to use camera-based system to finish this job like [8], [9]. FRAMEWORK
However, there performance can be negatively affected A. Problem Analysis
by low-light sleeping environment and can also raise (1) Mitigating the effect of people around. The Fresnel
privacy concerns. Zone model [1] is a quite reliable model to describe the
Recently, taking wireless radio signal as detection propagation of Wi-Fi radio signal in free space between a
media attract more and more attentions. Various RF-based pair of TX-RX antenna. This model accounts that in the
approaches ranging from Doppler radar [10], UWB radar context of radio propagation, Fresnel zones refers to the
[11] and FMCW radar [5] [7], to Universal Software concentric ellipses with foci in a pair of transceivers.
Radio Peripheral (USRP) based solutions [12], are quite When human respire in the Fresnel zone, the length of
accurate but cost too much and the devices are not widely reflection path changes with the ups and downs of chest. If
deployed in our daily living environments. the reflection pass length continuously changes by a
B. Wi-Fi based Respiration Detection wavelength λ, the resultant reflected signal will exhibit a
phase change (rotation) of 2п, and the phase rotation
As a drawback, the RF based systems mentioned above
reflected signal interfere with the Line of Sight (LOS)
rely on sophisticated and expensive hardware, making signal will lead to the amplitude changes. As the ups and
them impossible to large-scale deployment in ordinary downs of the chest have periodicity, the changes of
homes. In comparison commercial off-the-shelf (COTS) amplitude are periodic as well. This model can detect
Wi-Fi devices are low-cost and non-intrusive, which respiration precisely when there is only one person in the
builds a new way for convenient and inexpensive experimental environment; however, it is hard to detect
respiration detection. multi-user’s respiration as the detection usually is affected
In [13], [14] and [4], the received signal strength (RSS) by other users in the system.
measurements of a Wi-Fi based network of transceivers When we detect a user’s respiration in multiple user
are utilized for extracting respiration rate. The work [13] situations, with other people in the neighborhood, the total
and [14] is improved by [12] in which the RSS data from a receiving signal variance can be approximated as a
single TX-RX pair are able to track one’s respiration rate. combination of the variance caused by the activity of each
The RSS, however, has found to be insensitive for reliably person. The respiration information of the detected user is
tracking the minute chest movement due to the respiration, included in the shortest reflection path. In this way, we can
as the RSS changes caused by exhale and inhale are so mitigate the effect caused by other people around by
small that they can be easily submerged by environmental filtering out the data which is greatly affected by the
noise [3].
longer paths. In other word, we need to measure the
Compared with RSS, the Channel State Information
transfer time, based on which we can filter out the signal
(CSI) from Wi-Fi devices is richer and more sensitive for
coming from longer paths. For each user, we use TOA
detecting one’s respiration. In [1], they use CSI difference
caused by inhale and exhale to detected respiration even filter, therefore, we obtain the data without affection for
with the received CSI information they can figure out the each user.
posture changes during the person’s sleeping, however, In practice, CSI is measured on 30 selected Orthogonal
their work focus on respiration detection of a single user. Frequency Division Multiplexing (OFDM) subcarriers.
Zhang et al. [1] introduced the Fresnel Zone Model in OFDM technologies are implemented using a combination
free space, based on which they investigated the impact of of fast Fourier Transform (FFT) and inverse fast Fourier
human respiration on the receiving RF signals and Transform (IFFT) blocks. What we received is the radio
developed the theory to relate one’s breathing depth, RF signal in frequency domain, which cannot be directly
location and orientation to the detectivity of respiration. used to measure the transfer timer. However, with IFFT
However, their model requires an experiment environment processing of the CSI, we can obtain the channel response
without persons around the user, and if there are two in the time domain. Due to the bandwidth limitation, we
subjects’ respiration in their system they use Power cannot distinguish each signal path. However, by filtering
Spectral Density (PSD) to separate the respiration data of out the data whose TOA is bigger than a truncation
each person, but it is difficult to tell the breathing threshold, we can extract the information of LOS and the
information for different people. In other words, only one nearby NLOS path. In addition, due to the frequency-
person’s respiration can be detected at a time. selective fading, subcarriers are not affected equally by the
In this paper, we apply Fresnel Zone model to detect multipath effect. In other words, among the 30 subcarriers,
respiration, using multiple antenna pairs and the Time-of- some subcarriers are sensitive to multipath. For example,
according to [3], the No.1 to No.8 and No. 25 to No.30
subcarriers are more greatly affected by the multipath
 2017 IEEE 19th International Conference on e-Health Networking, Applications and Services (Healthcom)

effect. As a result, we select subcarriers which have a

Medium sensitivity to multipath, to help us minimize the
affective of person around.
(2) Increase density of Fresnel zone. According to the
Wi-Fi Fresnel Zone model, let ∆d be the distance of the
chest displacement, λ be the wavelength, then the phase
rotation θ caused by respiration can be calculated by the
following equation:

（1）
Figure 2. Respiration sensing with multiple Fresnel Zone
As the mean distance of chest movement ∆d in
(3) Verify if the quilt covered on the body will affect
anteroposterior for normal respiration and deep respiration
the respiration detect. Apart from the affect from the
are around 5mm and 12mm, respectively [2] [3], the
person around, there also might exist some covers (like
resultant phase change θ is between 60° to 150°.
quilt or clothes) on user’s body in daily life. When there
Meanwhile, there are some best places and worst places to
are some covers on user’s body, the ups and downs of their
sensing the respiration, according to [4]. While the worst
chest will become quite weak. For this issue, we have done
location for respiration sensing is around the boundary
a lot experiments to verify that the Wi-Fi RF signal can
within each Fresnel Zone, the best location appears in the
pass through the quilt without any affect, and covers
middle, as shown in Fig 1. To get the best accuracy, the
cannot affect respiration detection in our system.
simplest way is moving the devices to the best location
before detection. Obviously, such an approach is B. Overall Architecture
inconvenient to deploy. The overall architecture of our system is show in Fig.
To enhance the system’s ability to capture human 3 which is comprised of three steps. In the data collection
respiration, we use multiple TX-RX antenna pairs in this step, we use multiple TX-RX antenna to form a joint
paper. MIMO (Multiple-Input Multiple-Output), which is Fresnel Zone, and the density of Fresnel Zone is far
the basis of Multiple TX-RX, makes Wi-Fi devices greater than a single pair of TX-RX antenna. In the data
delivery signals using multiple transmitting and receiving processing step, we reduce the effect of surrounding
antennas. Each pair of transmit-receive antennas will form persons by removing the data whose TOA is bigger than a
their own Fresnel Zone respectively. With the existence of truncation threshold and select the subcarriers that are not
multiple Fresnel Zones, when a user locates on the greatly affected by multipath. Afterwards, we use the
boundary of one Fresnel Zone, she may locate on the wavelet filter and mean filter to filter out signal noise and
middle of another Fresnel Zone, as shown in Fig 2. By smooth the signal curve. In the respiration detection step,
comparing the peak to peak of each wave, we can select we use the peak-valley detection algorithm to detect peaks
the biggest one that is the most suitable TX-RX antennas of the wave, where each peak represents the occurrence of
pair to monitor the user’s respiration. one respiration.

IV. DETAILED DESIGN OF THE SYSTEM

A. Data Collection
We recruited one person to sit down between the
receiver and transmitter, and let him breath as usual.
When receiver received the RF signal mixed with the
reflected and LOS signal for 1 second, the receiver will
send the CSI information to our mini PC with MATLAB
for further processing.
With a single TX-RX antenna pair, we cannot always
detect human respiration. According to the Wi-Fi Fresnel
Zone model [2], while the worst location for respiration
sensing is around the boundary within each Fresnel Zone,
the best location appears in the middle, as shown in Fig 2.
Therefore, to get the best accuracy, we need to move
Figure 1. Respiration sensing in different location devices to the best location before detection. Obviously,
such an approach is inconvenient to deploy, so we use a
 2017 IEEE 19th International Conference on e-Health Networking, Applications and Services (Healthcom)

Figure 3. System architecture

MIMO receiver and a transmitter equipment to build a Specifically, the transmitter has 2 antennas and the
multiple TX-RX antenna system. receiver has 3 antennas, i.e., MT = 2 and MR = 3. We
implemented our experiment in a typical home
B. Data Processing
environment. We placed the receiver and transmitter on a
As our transmitter has 2 antennas and the receiver has pair of tripod with a distance of 100cm. The CSI values
3 antennas, therefore, we get 6 data streams in our system, are measured on ICMP ping packets sent from the Wi-Fi
and each stream includes the CSI information of 30 router to the laptop at the data rate of about 500 packets
subcarriers. According to [2], we choose one subcarrier per second.
which is most sensitive to respiration. As a result, we
finally get 6 CSI subcarries from 6 different data streams. B. Experiment Setting
Meanwhile, the original received CSI information is in Our experiment setting is shown in Fig.5, which
the frequency domain. Therefore, with IFFT processing of includes two receivers and one transmitter. Tow users sit
the CSI, we can obtain the channel response in the time down side by side, and there is a distance of 0.8 meter
domain. Moreover, by filtering out the data whose TOA is between them. The two receivers received the RF signal
bigger than a truncation threshold, we can extract the mixed with the LOA signal and the reflect signal from
information of LOS and the nearby NLOS path. two users. Obviously, the TOA of reflect signal from user
With the data filtered by the arriving time, we obtain A to the receiver B is longer than the TOA to receiver A.
the rough data without the effect form people around. In the counterpart, the TOA of reflect signal from user B
However, the voltage fluctuation caused by using to the receiver A is longer than the TOA to receiver B.
commodity devices is still unavoidable. The rough data We filter the reflected signal whose TOA is larger than a
still need the wavelet filter and mean filter to smooth RF truncation threshold for each receiver.
signal. In the end, we get 6 smoothing data streams.
C. Respiration Detection
As mentioned before, different location in Fresnel
Zone has different sensing capability on respiration. After
we get 6 processed data streams from 6 different antenna
pairs, we need to figure out which antenna pair is the best
one to detect respiration.
After the detection of the peak-valley, we further use a
sliding window about 0.1 second to figure out the stream
with the biggest peak valley difference. In other words,
this data stream fits the present chest location best.
Obviously, we should choice this data stream to detect
Figure 4. Experiment Setting
respiration.
V. EXPERIMENTAL EVALUATION C. Result of Multiple Users’ Respiration Detection
Using Multiple Antenna Pairs
A. Hardware Setup
As we mentioned before, with multiple antenna pairs,
We used a laptop with Intel Link 5300 Wi-Fi NIC as each antenna pair can detect respiration independently.
the receiver, which has 4GB of RAM and Ubuntu 14.04 However, there is a pair of antennas has the best position
as its operating system. TP-Link TLWR1043ND Wi-Fi to detect respiration, as show in Fig.5 (a). The 6 lines of
router was adopted as the transmitter operating in 802.11n different colors represent the 6 processed data streams
AP mode at 2.4GHz. We use a mini PC with Ubuntu as its come from 6 different pair of antennas. Obviously, at
operating system and the MATLAB for data processing.
 2017 IEEE 19th International Conference on e-Health Networking, Applications and Services (Healthcom)

(a) (b)
Figure 5. Respiration detection result: (a) people A, (b) people B

least 5 of them have detected the respiration data, when the person respires the quilt ups and downs
however, there is only one stream (i.e., the one at the obviously, and the signal we obtained has no obvious
bottom of the 6 lines) has the best performance to monitor difference with the signal obtained without a quilt covered
the respiration. Obviously, with the TOA filter, the over the chest.
respiration wave of each user was captured clearly. The second group of experiments we have down is to
cover a thick quilt over one person’s chest. In this
D. The Result of Mitigating the Effect of People Around situation, when the person respires, the chest below the
In the control group, without TinySense, we detected quilt still ups and downs, however, as the quilt is too thick,
the respiration rate of only one subject without any person there is some space between the chest and the quilt, the
around for about 5 minutes, and the respiration rate of one chest and the quilt do not move ups and downs in the
subject with another person moving around him. Then we same rhythm.
detected the respiration in the same condition with our Nevertheless, according the data we obtained, we still
system, the result is show in Fig.6. When there is only one can detect respiration under the thick quilt as show in
person, the accuracy is over 95% not matter using or not Fig.7. The RF signal can pass through the thick quilt to
using TinySense. However, when there is some person detect the respiration. As a result, covering quilt does not
around the user, the accuracy of using TinySense is over severely impact the performance of respiration detection.
88%, while the performance of not using TinySense
declines to 70%.

Figure 7. Respiration detect result with/without quilt.

Figure 6. Accuracy of respiration rate VI. DISCUSSIONS

We have reported our preliminary efforts on human
respiration detection with Wi-Fi signals in this work. In
E. Verify Whether the Quilt Affect the Respiration this section, we will discuss several limitations and
Detection opportunities to improve our work in the future.
To verify this issue, we have done two groups of According to the experimental results, we can achieve
experiments. In the first group, we cover a thin quilt over more than 88% accuracy when detecting the respiration of
one person. The quilt is quite close to the chest, therefore, 2 users at a time. Though, the accuracy might be affected
 2017 IEEE 19th International Conference on e-Health Networking, Applications and Services (Healthcom)

by several factors, such as the distance between the two [2]. J. Wang, et al. LiFS: low human-effort, device-free localization
subjects and the distance between the receivers, the high with fine-grained subcarrier information. In Proceedings of the
22nd Annual International Conference on Mobile Computing and
of the receivers we set, and so on. Furthermore, Wi-Fi Networking (MobiCom’16), 2016, pp. 243-256.
signals are not that stable and can be affected by many [3]. X. Liu, et al. "Contactless Respiration Monitoring Via Off-the-
factors, such as the voltage fluctuation. As an early effort, Shelf WiFi Devices." IEEE Transactions on Mobile Computing,
we make some restrictions in the experiments. In the vol. 15 no.10, 2016, pp. 2466-2479.
future, we plan to make more experiments to study the [4]. N. Patwari, et al. "Breathfinding: A Wireless Network That
detailed impacts of such factors and improve our Monitors and Locates Breathing in a Home." IEEE Journal of
approach to make it more practical for smart home usage. Selected Topics in Signal Processing, vol.8 no.1, 2014, pp. 30-42.
Specifically, TinySense need a special layout of [5]. F. Adib, H. Mao, Z. Kabelac, D. Katabi, and R. C. Miller. Smart
homes that monitor breathing and heart rate. In Proceedings of the
receivers and transmitter. In case of detecting more than 33rd Annual ACM Conference on Human Factors in Computing
two persons, such as four people, the layout should be as Systems. ACM, 2015, pp. 837–846.
Fig. 8. In the future, we will try to reduce the number of [6]. Tanita. Available: https://www.tanita.co.jp/products/models/
receivers, explore its effective bounds, and improve the sl501.html/, 2017.
human respiration detection performance. [7]. F. Adib, Z. Kabelac, H. Mao, D. Katabi, and R. C. Miller, “Demo:
Real-time breath monitoring using wireless signals,” in Proc. 20th
Annu. Int. Conf. Mobile Comput. Netw., 2014, pp. 261–262.
[8]. J. Paalasmaa, M. Waris et al., “Unobtrusive online monitoring of
sleep at home,” in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol.
Soc., 2012, pp. 3784–3788.
[9]. Philips vital signs camera. Available: http://
www.vitalsignscamera.com/, 2017.
[10]. H Erhan Dincer and William O’Neill. Deleterious effects of sleep-
disordered breathing on the heart and vascular system.
Respiration, Respiration, vol. 73, no. 1, 2006, pp. 124–130.
[11]. NHLBI: Health Information for the Public. U.S. Department of
Health and Human Services. Sleep apnea: What is sleep apnea?
2010.
Figure 8. Layout of four persons
[12]. GigaOm. 2016. Could a breath-monitoring headset improve your
VII. CONCLUSION health?https://gigaom.com/2013/09/20/could-a-breathmonitoring-
headset-improve-your-health/, 2016..
In this paper, we have proposed a solution for [13]. N. Patwari, J. Wilson, S. Ananthanarayanan, S. K. Kasera, and D.
respiration detection for multiple users. We use two steps R. Westenskow, “Monitoring breathing via signal strength in
to reduce the effect caused by other people around. The wireless networks,” IEEE Trans. Mobile Comput., vol. 13, no. 8, ,
first step is filtering out the data whose TOA is bigger Aug. 2014, pp. 1774–1786.
than a truncation threshold. The second step is selecting [14]. C. Scully, L. Jinseok, J. Meyer, A. M. Gorbach, D. Granquist-
Fraser, Y. Mendelson, and K. H. Chon, “Physiological parameter
the subcarriers that are not greatly affected by multi-path. monitoring from optical recordings with a mobile phone,” IEEE
In addition, we use multiple TX-RX antenna pairs to Trans. Biomed. Eng., vol. 59, no. 2, 2012, pp. 303–306.
enhance the ability of our system to capture human [15]. O. J. Kaltiokallio et al., “Non-invasive respiration rate monitoring
respiration, which can expand the distribution of the using a single COTS TX-RX pair,” in Proc. 13th Int. Symp. Inf.
Fresnel Zone. Experimental results show that our Process. Sens. Netw., 2014, pp. 59–70.
approach can efficiently detect the respiration of multiple [16]. H. Abdelnasser, K. A. Harras, and M. Youssef, “UbiBreathe: A
users. We will in the future explore the usage of our ubiquitous non-invasive WiFi-based breathing estimator,” in Proc.
16th ACM Int. Symp. Mobile Ad Hoc Netw. Comput., 2015, pp.
model for other tiny sensing of human behaviors, such as 277–286.
human identification [20] and gesture recognition. We
[17]. Okeyo G., Chen L., Wang H., and Sterritt R., Ontology-based
will further investigate simple deployment of our system Learning Framework for Activity Assistance in an Adaptive
and its usage in different environments. Smart Home, in book “Activity Recognition in Pervasive
VIII. AckNOWLEDGMENT Intelligent Environments”, Atlantis Ambient and Pervasive
Intelligence, Vol.4, ISBN:9789078677352, Atlantis Press, 2011
This work was partially supported by the National pp.237-262.
Key R&D Program of China(2017YFB1001800)，and [18]. Poland M.P., Nugent C.D., Wang H., Chen L., Genetic Algorithm
and Pure Random Search for Exosensor Distribution Optimisation,
the National Natural Science Foundation of China (No. International Journal of Bio-Inspired Computation, Vol. 4, No. 6,
61332005, 61373119, 61402369,61772428) October 2012, pp.359–372.
REFERENCES [19]. Skillen K., Chen L., Nugent C.D. and Donnelly M., “A User
Profile Ontology Based Approach for Assisting People with
Dementia in Mobile Environments”, The 34th Annual
[1]. H. Wang, et al. Human respiration detection with commodity wifi International Conference of the IEEE Engineering in Medicine
devices: do user location and body orientation matter?. In and Biology Society (IEEE EMBC 2012), 2012, pp. 6390 - 6393.
Proceedings of the 2016 ACM International Joint Conference on
[20]. X. Tong, Guo B., Wang Z., Li M., Yu Z. "FreeSense: Indoor
Pervasive and Ubiquitous Computing (UbiComp’16). ACM, New
Human Identification with WiFi Signals", The IEEE Global
York, 2016, pp. 5-36.
Communications Conference (IEEE GlobeCom'16), 2016.