IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO.

11, NOVEMBER 2015 2329

Non-Invasive Detection of Moving and

Stationary Human With WiFi
Chenshu Wu, Student Member, IEEE, Zheng Yang, Member, IEEE, Zimu Zhou, Student Member, IEEE,
Xuefeng Liu, Member, IEEE, Yunhao Liu, Fellow, IEEE, and Jiannong Cao, Fellow, IEEE

AbstractNon-invasive human sensing based on radio signals Typical applications include human detection for intruder de-
has attracted a great deal of research interest and fostered a tection, emergency responses, and in-home children and elderly
broad range of innovative applications of localization, gesture monitoring. Pioneer efforts have explored the possibility of ex-
recognition, smart health-care, etc., for which a primary primitive
is to detect human presence. Previous works have studied the tracting motion information from wireless signals to localize or
detection of moving humans via signal variations caused by human track whole-body motions [1][3] or even gestures [4], [5] non-
movements. For stationary people, however, existing approaches invasively. The term non-invasive, a.k.a passive or device-
often employ a prerequisite scenario-tailored calibration of chan- free, means that people are not assumed to carry any wireless
nel profile in human-free environments. Based on in-depth under- device. The primary underpinning is that a persons movement
standing of human motion induced signal attenuation reflected by
PHY layer channel state information (CSI), we propose DeMan, a can modulate wireless signals and result in temporal changes
unified scheme for non-invasive detection of moving and stationary that are observable from received signals [1]. The mobility of
human on commodity WiFi devices. DeMan takes advantage of to-be-observed users makes a prerequisite of existing passive
both amplitude and phase information of CSI to detect moving human detection systems. However it is intrinsically challeng-
targets. In addition, DeMan considers human breathing as an ing to detect stationary users based on radio reflections [6].
intrinsic indicator of stationary human presence and adopts so-
phisticated mechanisms to detect particular signal patterns caused To detect the presence of stationary people, existing schemes
by minute chest motions, which could be destroyed by significant employ a prerequisite calibration of channel profile in human-
whole-body motion or hidden by environmental noises. By doing free environments, and simplify human presence as shadowing
this, DeMan is capable of simultaneously detecting moving and on the Line-Of-Sight (LOS) path [7]. The calibration needs to
stationary people with only a small number of prior measurements be conducted offline to collect a link profile for human-free
for model parameter determination, yet without the cumbersome
scenario-specific calibration. Extensive experimental evaluation settings and online detection is accomplished by comparing the
in typical indoor environments validates the great performance real-time measurements against the static profile. Despite the
of DeMan in various human poses and locations and diverse cumbersome scenario-tailored profiling, such schemes may still
channel conditions. Particularly, DeMan provides a detection rate fail due to temporal environmental unstableness and multipath
of around 95% for both moving and stationary people, while effects [8].
identifies human-free scenarios by 96%, all of which outperforms
existing methods by about 30%. In this work, we ask the following questions: Is it possible
to detect stationary people passively without any scenario-
Index TermsNon-invasive, human detection, calibration-free, tailored calibration? Furthermore, can we build a unified
human breathing, channel state information.
framework with commodity WiFi devices to simultaneously
I. I NTRODUCTION detect both moving and static persons? We investigate the inter-
ference of human presence on wireless signals and demonstrate

R ECENT advances in WiFi techniques have enabled a

range of ubiquitous applications where wireless signals
convey human body induced radio shadowing and reflections.
that the rhythmic chests rise and fall that alternate between
inhaling and exhaling of human breathing induce repetitive
changes on received signals, shedding promising lights on
the non-invasive detection of static people. To deliver these
Manuscript received August 15, 2014; revised December 7, 2014; accepted observations into a practical human detection scheme, multiple
January 29, 2015. Date of publication May 6, 2015; date of current version challenges need to be overcome: 1) How to discover and
October 19, 2015. This work is supported in part by the NSFC Major Program harness weak human breathing patterns from wireless signals
61190110, the NSFC under grants 61171067, 61472098, and 61361166009,
and the NSFC Distinguished Young Scholars Program under Grant 61125202. in the presence of environmental unstableness? 2) How to
The research of X. Liu and J. Cao is supported in part under NSFC/RGC Joint integrate the detection of moving and stationary humans into
Research Schemes 3-ZG1L and the NSFC under Grant No. 61332004. a unified framework without obscuring the minute breathing-
C. Wu, Z. Yang, and Y. Liu are with the School of Software and TNList,
Tsinghua University, Beijing 100084, China (e-mail: [email protected]; induced chest motion with significant body motions?
[email protected]; [email protected]). To address these challenges, we propose DeMan, a unified
Z. Zhou is with the Hong Kong University of Science and Technology, scheme for non-invasive DEtection of moving and station-
Kowloon, Hong Kong (e-mail: [email protected]).
X. Liu and J. Cao are with the Department of Computing, Hong Kong Poly- ary huMAN with commodity WiFi devices. DeMan leverages
technic University, Kowloon, Hong Kong (e-mail: [email protected]; the PHY layer Channel State Information (CSI) provided by
[email protected]). commercial WiFi products, which offers fine-grained channel
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. description at the granularity of OFDM subcarriers [9]. To
Digital Object Identifier 10.1109/JSAC.2015.2430294 differentiate macro human movements and micro chest motion
0733-8716 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2330 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 11, NOVEMBER 2015

of breathing, we propose a motion interference indicator based The rest of the paper is organized as follows. We review the
on the variances of CSI to provide a primary judgement of two related works in Section II and present some preliminaries in
cases: 1) if a moving person is more likely to present, DeMan Section III. Section IV presents an overview of the system as
starts the moving human detection module; 2) otherwise the well as designs of the motion interference indicator and moving
stationary human detection module is chosen. target detection, while details of stationary target detection are
On one hand, to detect moving targets, we explore the introduced in Section V. Section VI provides the performance
full potential of CSI in both perspectives of amplitude and evaluation. We discuss the limitations and future works in
phase. We demonstrate that phase information is similarly or Section VII and conclude this work in Section VIII.
even more sensitive to environmental changes, which, however,
has not been sufficiently utilized to the comparable extent of
II. R ELATED W ORKS
amplitude [10]. To extract environment-independent features
of both signal amplitude and phase, we propose the maximum The design of DeMan is closely related to the following
eigenvalues of correlation matrices of successive measurements categories of research.
to characterize the variations of temporal wireless signals. Wireless Non-Invasive Human Detection: Wireless non-
On the other hand, to detect stationary people, DeMan pro- invasive human detection systems detect and localize humans
cesses the received signals with a bandpass filter to extract via their impact on received signals, while the targets carry no
the signal components within an interest frequency band (cor- wireless-enabled devices. The basic principles differ for station-
responding to the normal frequency range of human breath- ary and mobile targets. To detect moving users, the variance
ing). Then we justify a sinusoidal model to formulate the of the RSS measurements is directly compared with a pre-
breathing-induced wave-like patterns on wireless signals and defined threshold [1], while the mean of RSS measurements is
detect a breathing person by searching for periodic signals of compared with that of a normal profile when there are no users
concerned frequencies. DeMan further harnesses the frequency in the monitored area to detect motionless users [7]. Recent
diversity of modern OFDM modulation to enable static people advances explore stationary and moving target detection based
detection under complicated and diverse human poses and on signal envelope features [6] at the cost of a dense deployed
locations. wireless links. Some works develop sophisticated mechanisms
We prototype DeMan on commodity WiFi devices and eval- for through-wall imaging of subjects (including occluded ones)
uate its performance in various scenarios of typical indoor using RF signals [11][13]. One of the latest innovations [14]
buildings. Experiment results, from over 8-hour measurements develops a theoretical and experimental framework with only
in one week, demonstrate that DeMan achieves respective WiFi power measurements for the problem. In this work, we
true positive rate for moving and stationary people of 94.82% also aim at detecting both stationary and moving humans, yet
and 93.33% and a true negative rate of 96.25% for human- dive into the PHY layer in purpose of achieving robust detection
free scenarios, outperforming previous approaches by about with a single wireless link in even multipath-dense indoor
30%. DeMan uses a small number of prior measurements to environments. Our scheme is able to distinguish the impact of
determine several scenario-independent parameters, which are a stationary human and static environmental interference such
then applicable to different contexts. Consequently, DeMan as the location change of furniture by capturing the unique
requires no scenario-specific calibration, which is beyond the breathing patterns.
achievement of previous works for stationary target detection. Wireless-Based Gesture and Activity Detection: Since wire-
We envision it as an important step towards fully practical less signals may be reflected differently with changes of human
technology of device-free human detection. postures, numerous efforts have utilized wireless signals to
In summary, the core contributions are as follows: detect whole-body [4], [15] or hand gestures [16] and daily
activities [17] by analyzing the received signal patterns. Some
We propose a unified framework for simultaneous detec- work extracts Doppler features from received signals using
tion of moving and stationary people. To the best of our customized OFDM signal processing [4], or leveraging In-
knowledge, this is the first solution that converges the verse Synthetic Aperture Radar (ISAR) to enable through-wall
advantages of purely WiFi-based, scenario-specific cal- gesture sensing [15]. Alternatively, a pattern matching based
ibration free, and non-invasive together in the literature. approach can be employed to recognize hand gestures [5] or
We design and implement a unified detection approach daily activities [17] on commodity WiFi devices. The tradeoff,
for stationary persons by modeling and exploiting minute however, is to build up a gesture profile database in-advance.
chest motions of human breathing as an intrinsic indi- In this work, we also aim to detect humans via their reflected
cator for human presence. Different from power fading, signal patterns, but target at the much more micro motion, i.e.,
chest motion analysis enables DeMan to accurately de- breathing. We try to capture the tiny impact of breathing on
tect stationary people not only on the direct LOS path, wireless signals harnessing the repetitive patterns of breathing.
but also on the reflected paths with a single wireless link, Contactless Breath Detection: Breath is an important vital
resulting in an extended sensing coverage. sign and active research has been conducted to monitor breath
We investigate previously unexplored phase information via chest movements or inhaling airflow measured by wearable
of CSI and propose a novel method to extract and analyze sensors [18]. A promising alternative is to exploit wireless
phase feature, which is demonstrated to improve the signals to detect breathing unobtrusively by capturing chest mo-
accuracy and sensitivity of moving target detection. tions during breathing [19] utilizing costly radar infrastructure.
WU et al.: NON-INVASIVE DETECTION OF MOVING AND STATIONARY HUMAN WITH WiFi 2331

Some pioneer work has demonstrated the viability of detecting

breaths using commodity wireless devices [20], yet requires
multiple transceivers to create a dense network of links. The
closest to our works are [21] and [22], which enable non-
intrusive breath monitoring on a single wireless link with di-
rectional antennas and dedicatedly placed transmitter-receiver
pairs, respectively. Nevertheless, they are effective only when
people present closely to short LOS links. Moreover, these
systems are primarily designed for health-care and sleep mon- Fig. 1. System architecture of DeMan.
itoring applications, where persons are required to be lying
on beds. Hence, they cannot be directly adopted for intruder Compared with the conventional MAC layer Received Signal
detection or other ubiquitous applications, where users are not Strength Indicator (RSSI), CSI offers finer-grained information
expected to always appear at a fixed location. Unlike [20][22], in two aspects:
DeMan is able to detect breaths over a wider area without Phase Information: In multipath-dense indoor environ-
dense links and designed link placement, thus extending its ments, wireless signals usually propagate to the receiver
applications to intruder detection, emergency responses, and in- through multiple paths. These multipath components can
home children and elder monitoring, etc, where user presence superpose either constructively or destructively depend-
is unrestricted and unpredictable. ing on their relative phases. Since RSSI only offers am-
CSI-Based Indoor Localization: As a promising substitute plitude information, previous research explores indirect
for MAC layer RSSI, the fine-grained PHY layer CSI available proxy such as fade level to infer the phase superposition
on commercial WiFi devices has raised increasing enthusiasm status [8]. In contrast, CSI provides both subcarrier phase
on CSI based indoor localization with meter-level accuracy and amplitude information, which holds potential for
[23], [24]. Since CSI depicts frequency diversity at the gran- more sensitive and accurate motion detection.
ularity of OFDM subcarriers, it also benefits non-invasive Frequency Diversity: In essence, RSSI depicts the total
human detection as a more informative signature, and has received power across all subcarriers. Therefore, RSSI
been employed in non-invasive motion detection [25], entity fails to characterize multipath propagation, which may
localization [23], crowd counting [26] and walking activity convey the subtle breathing patterns. Since the subcar-
recognition [17]. Our human detection system also builds upon riers in OFDM tend to fade independently, CSI brings
the fine-grained CSI measurements. Nevertheless, unlike ex- about opportunities to optimize and magnify the breath-
isting efforts that either assume moving targets [25], [26] or ing patterns leveraging frequency diversity. It may also
require signature collection and matching [17], we provide a extend detection range since breathing patterns from a
unified framework to detect both moving and stationary humans NLOS path may also be captured and resolved.
without prior signature collection.
In a nutshell, CSI exposes a finer-grained spectral structure
of wireless channels. In the subsequent sections, we strive
III. P RELIMINARY to harness the subcarrier phase and amplitude information to
design a unified framework for both moving and stationary
With the increase of operating bandwidth and the need human detection.
to support MIMO techniques, current WiFi devices start to
track fine-grained channel measurements leveraging Orthogo- IV. D E M AN D ESIGN
nal Frequency Division Modulation (OFDM). An OFDM WiFi
transmitter (IEEE 802.11a/g/n) sends bits through multiple This section presents the design of DeMan in a top-down
subcarriers in parallel, and the receiver detects the start of each manner, with emphasis on moving target detection. As a uni-
OFDM packet via a pre-defined preamble [27]. The preamble fied framework, DeMan simultaneously detects both moving
also facilitates the receiver to estimate channel conditions on and stationary targets in a generic processing flow without
each subcarrier [28]. Channel condition on each subcarrier scenario-specific pre-training. Yet due to the unique challenges
involves both amplitude attenuation and phase shift, and can involved in stationary target detection without scenario-specific
be represented by a complex number. With commodity WiFi calibration, we defer the details on stationary target detection in
Network Interface Cards (NICs) such as Intel 5300 and slight Section V. Here a target refers to a person only and we use the
firmware modification, a group of 30 subcarrier channel mea- two words interchangeably henceforth.
surements can be revealed to upper layer users in the format of
Channel State Information (CSI) [9]: A. Overview

 Hk The architecture of DeMan mainly consists of two compo-

Hk = Hk ej (1) nents: moving target detection and stationary target detection.
As shown in Fig. 1, DeMan works as follows.
where Hk is the CSI at the subcarrier k with central frequency DeMan initializes by extracting CSI compatible with IEEE
of wk , and Hk  and  Hk denote its amplitude and phase, 802.11n standards on commodity NICs. The raw CSI mea-
respectively. surements are passed through the Hampel identifier [29] to sift
2332 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 11, NOVEMBER 2015

Fig. 3. Human movements induce significantly larger changes in signal

Fig. 2. An illustration of impacts of human presence on signal propagation. envelope.
(a) Static environment without human. (b) Moving human continuously induces
significant changes on RF signals. (c) Breathing motion on the direct LOS
path incurs different extent of shadowing. (d) Breathing motion shifts signal
propagation between LOS and NLOS paths. (e) Breathing motion alters a
reflection path that is generated by the human body. (f) Breathing motion
modulates an existing reflection path.

outlier observations [17]. Afterwards the CSIs are processed by

a lightweight motion indicator to coarsely decide whether there
is a moving person. If the answer is YES, the CSI measure-
ments are further fed into the moving target detection module
for finer-grained motion detection. Otherwise, the stationary
target detection module is triggered to identify static human
presence. DeMan reports a detected event if either module
outputs affirmatives. Otherwise, no person is detected within
the monitoring area. Fig. 4. Envelope variances of human movements are remarkably larger than
To reliably detect moving humans, we exploit both amplitude those of no movements.
and phase information conveyed in CSI measurements. We
calculate the correlation matrices for the complex CSI measure-
ments for each subcarrier, and extract the corresponding max- original propagation paths (either LOS or reflection paths). In
imum eigenvalues for each correlation matrix. The maximum contrast, as shown by other figures in Fig. 2, the propagation
eigenvalues are then combined into a two-dimensional feature paths experience subtle changes in case of stationary persons or
to infer the presence of moving people. no person presence.
To detect stationary persons without scenario-tailored cal- To validate these observations, we first conduct primary
ibration (profiling and comparing with signal characteristics measurements in case of moving human, stationary human,
collected in a human-free environment for each case), we and no human respectively, with a commercial wireless router
harness the observation that humans, even when standing still as transceiver and a laptop as receiver. Since we intend to
or remaining seated, exhibit unique motion patterns like distinguish the case of mobile people from other cases, data
breathing. We first filter the CSI measurements to suppress collected in presence of static human and no human are in-
signals irrelevant to breathing frequencies. The filtered CSI tegrated into one category marked as no movements here.
sequences are then fitted with a sinusoidal breathing model to Fig. 3 experimentally demonstrates the signal strength fluctu-
estimate its dominant frequency. If the estimated frequency falls ations of measurements with and without human movements,
within the frequencies of normal human breaths, the presence of each for 5 seconds. As is shown, human movements can induce
a person is detected. Otherwise, the stationary target detection changes of up to 4 dB in signal envelope, while changes
module announces negative. caused by electronic noise, quantization errors and/or human
breathing in stationary scenarios are mostly limited within
only 1.5 dB. In other words, CSI exhibits remarkably larger
B. Motion Interference Indicator
variations in case of moving humans than in case of static
It is widely recognized that RF signals would fluctuate humans or no human, although CSI is also sensitive to breath-
remarkably when objects move within the area of interests, ing motion of static humans as we demonstrate later. Denote
and remain stable in case of no motion interference [1], [25]. the mean-removed signal strength as X. Then X is expected
The motion interference indicator exploits such motion-induced to follow a zero-mean Gaussian X N (0, 2 ). Fusing more
signal fluctuation to infer potential moving persons. data together, Fig. 4 further consolidates the observation that
As shown in Fig. 2(b), when a person passes through a human motion induces consistently larger variances 2 than
wireless link, his/her presence will continuously violates the static cases.
WU et al.: NON-INVASIVE DETECTION OF MOVING AND STATIONARY HUMAN WITH WiFi 2333

Motivated by these observations, we build a lightweight

motion interference indicator based on the variance of signal
envelope using a hypothesis testing:

Sm : 2 > th(m)
2
(2)
Ss : 2 < th(s)
2

where Sm and Ss indicate the state of motion and static case, re-
spectively. th(m)
2 and th(s)
2 denote the corresponding threshold
to trigger moving target detection and stationary target detec-
tion. Recall Fig. 4, there are a small portion of cases where the
variances of motion and motionless cases are similar. Noticing
this, we set th(m)
2 < th(s)
2 so that a critical zone always exists as
[th(m), th(s)]. To avoid false decision in the motion indication
2 2 Fig. 5. Normalized maximum eigenvalues of amplitude and phase correlation
matrices.
stage, those confused cases fallen in the critical zone are doubly
checked by both subsequent modules. The thresholds can be
determined by some preliminary measurements. Different from where and are intuitively the slope and offset of phase
previous works that rely on scenario-tailored calibration, how- change over all the subcarriers, respectively. Then the sanitized
ever, we do not need to calibrate for each different case because phase measurements are re-assembled with the corresponding
the thresholds can gracefully apply to various environments. amplitudes into complex CSIs for mobile target detection.
Moreover, we design a critical zone of the thresholds to tolerate Consider N CSIs within a time window T, where each CSI on
a range of potentially different values in diverse scenarios. subcarrier k sampled at time ti , i [1, N] is a complex number
The design of the motion indicator is lightweight since the as in Equation (1):
calculation of envelope is fast and effective. The hypothesis  Hk (i)
testing provides a primary indication of motion or motionless, Hk (i) = Hk (i) ej , (3)
but not the ultimate declaration with high confidence. Actually,
while the usage of CSI variations is sufficient for motion where  Hk (i) denotes the revised phase k (i). Then CSI sam-
indication, it is too optimistic to be an effective metric for target ples at time ti over all the n subcarriers form a complex vector
detection since it is too sensitive to environment dynamics.
H(i) = [H1 (i), H2 (i), , Hn (i)] . (4)
A more elaborative metric is proposed for moving human
detection in the following section. Since human motions induce temporal fluctuations of the
received signals, we investigate to depict such temporal dis-
C. Moving Target Detection turbance via correlations between successive measurements.
Concretely, for the N CSIs H = [H(i)]Nn , we calculate the
If a group of measurements is labelled with state Sm in the respective correlation matrices A and C for amplitudes and
previous stage, it is then passed through a more elaborate and phases of the CSI measurements as follows:
reliable examination, i.e., moving target detection.    
Numerous research has explored to detect human movements A = a(i, j) NN , C = c(i, j) NN , (5)
non-invasively for localization and tracking [1], [3], counting
[2], [26] or activity recognition [17]. However, most utilize only where each element denotes the correlation coefficient between
the envelope features of received signals, either in the form of H(i) and H(j):
value of MAC layer RSSI [1], [2] or amplitude of PHY layer
CSI [17], [26], yet ignore the counterpart phase information. In a(i, j) = corr (H(i) , H(j)) , (6)
this work, we demonstrate that phase information is similarly c(i, j) = corr ( H(i),  H(j)) . (7)
or even more sensitive to environmental changes and thus
provides more accurate detection of moving humans. Hence we Afterwards, we derive the eigenvectors of matrices A and C
incorporate both amplitude and phase features to unleash the and exert the normalized maximum eigenvalues (denoted as
full potential of CSI for more accurate and reliable detection. A and C for A and C respectively) for moving human de-
Due to the lack of time and frequency synchronization, how- tection. Generally, in case of no human presence or merely
ever, raw phase information extracted from commodity WiFi stationary human, successive measurements would exhibit high
devices tends to be extremely random [24], [26]. Denote k = correlation factor, resulting in large eigenvalues (close to 1). In
k + 2wk t + 2wk w as the measured phase at subcarrier contrast, the eigenvalues tend to be small if there are moving
k with carrier frequency wk , where k is the genuine phase. humans during the measurements. Fig. 5 depicts an illustration
2wk t and 2wk w are the unknown phase shifts caused by of 500 groups of measurements for each case, of which each
the clock offset t and frequency difference w. group involves 500 packets. As seen, both A and C are close
To mitigate the random noise in raw phase measurements, to 1 in case of stationary human presence or no human presence,
we employ a linear transformation as recommended in [24]. while decrease dramatically in case of human movements. Con-
We revise the raw phase information as k = k wk sequently, one can easily search for a cutting edge (threshold) to
2334 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 11, NOVEMBER 2015

Fig. 6. Rhythmic motions of human breathing induce wave-like patterns on the received signals. Various colors in (a) and (c) indicate different level of signal
strengths, decreasing from red to blue. (a) Raw CSI measurements with human breathing for 1 minute. (b) periodic CSI patterns at subcarrier level (Subcarrier #10).
(c) conspicuous repetitive patterns over all subcarriers of filtered CSIs.

identify measurements accompanied with human movements. In typical indoor environments, wireless signals can prop-
In addition, the threshold is scenario-independent thanks to agate via reflection, diffraction, scattering via human bodies
the use of eigenvalue-based features, which are independent and other environmental obstacles. Therefore, signals can be
of absolute signal powers that vary over different scenarios potentially modulated by periodic chest motions of breathing
and different time. This threshold, together with the ones for if it interacts with the person, even when he stands still at the
motion interference in Equation (2), are the only components same place.
of DeMan that need slight prior efforts to calibrate, except for As illustrated in Fig. 2, when a person is present on the LOS
which DeMan requires no calibration. path, the movements of chest cavity would either dictates the
The use of eigenvalues of a correlation matrix is indepen- signal propagation by different extent of shadowing (Fig. 2(c)),
dent of absolute signal powers. Therefore once the prerequi- or continuously shifts the propagation path between LOS and
site thresholds are determined, they are applicable to various NLOS (Fig. 2(d)). Several works [20], [21] have observed sim-
scenarios and do not need to be re-calibrated. This approach ilar phenomenon on RSSI as Fig. 2(d) and deploy dedicatedly
is inspired by [25] that employs maximum and the second placed TX and RX for breath monitoring. Nevertheless, they
maximum eigenvalues of amplitude, and we advance it by are effective only when the conditions in Fig. 2(d) are strictly
harnessing both amplitude and phase features of CSI. satisfied.
In presence of a breathing person off the LOS path, the re-
ceived signals can also be continuously disturbed by reflections
V. S TATIONARY TARGET D ETECTION
from the moving chest. As shown in Fig. 2(e), the presence
In contrast to motion detection, conventional solutions to of a human body would create a new reflection path while
device-free detection for stationary humans often require a prior the persons breathing can repeatedly change that path. Also, a
profile measured with no human presence within the monitored breathing person could refashion an already existing reflection
area [7]. This is because the presence of stationary human path, as illustrated in Fig. 2(f). Such changes in multipath prop-
normally incurs static signal strength changes by shadowing the agations, however, are scarcely possible to be captured by the
LOS path or creating a new reflection path [8], but only slight coarse-grained RSSI, yet observable through CSI, which has
fluctuation within a temporal window. Hence prior awareness been demonstrated to be capable of characterizing multipath
of the signal profiles without human presence is indispensable effects at OFDM subcarrier level [10].
as reference for individual cases. To eliminate the overhead of Measuring Human Breathing: Although human breathing
such pre-calibration and achieve a unified detection framework does change signal propagation, are the alterations discernible
for both moving and stationary targets, the key observation and measurable using commodity WiFi infrastructure? We give
is that stationary people continuously breathe, which can be a positive answer to this critical question after conducting real
detected from wireless signals if elaborate mechanisms are experimental measurements using commodity WiFi devices.
designed for subtle chest motions. As a preliminary verification, we collect a group of mea-
surements by letting one volunteer stand still and breathe
naturally beside a commercial laptop that serves as the receiver.
A. Periodic Alterations From Breathing
The measurements last for 2 minutes and the original CSI
We begin with some intuitive observation and formal justifi- amplitudes are plotted in Fig. 6(a). The amplitude patterns
cation on why human breathing is measurable via CSI. are different from those of completely human-free static en-
Analyzing Human Breathing: Traditional non-invasive hu- vironments. By employing a bandpass filter (details will be
man detection schemes mostly detect whole-body human mo- discussed in the following section) on the original CSIs, we
tions [7], and assume that the signals remain nearly constant in can see more conspicuous periodically oscillatory patterns, as
static environments. We demonstrate that, however, the wireless shown in Fig. 6(c). Similar periodicities arise over almost all
signals are sensitive enough to be distracted by breathing people subcarriers, since breathing produce consistent interference on
who stands still on, close to, or distant to the LOS path. all subcarriers. Thus viewing from individual subcarrier, one
WU et al.: NON-INVASIVE DETECTION OF MOVING AND STATIONARY HUMAN WITH WiFi 2335

2) Sinusoidal Model: Recall Fig. 6(b), breath-induced signal

presents a sinusoid-like pattern. Previous works that moni-
tor breaths with UWB or sensor networks also suppose that
breathing attaches an additional sinusoidal component to the
received signal [20], [21]. Intuitively, this is because periodic
chest motion produces sinusoidal time delays of the signals
reflected by chest [20]. For the sake of simplicity, we still use
Hk (i) to denote the amplitude of CSI on subcarrier k measured
at time ti in this section. In a static environment, the amplitude
of received signals can be expressed as

Hk (i) = Hk + k (i), (8)

Fig. 7. Power spectral density of the filtered signal.
where Hk is the mean amplitude of the received signal at
subcarrier k and k (i) is an additive noise. When a breathing
can observe obvious wave-like patterns arising in the filtered person presents, an additional sinusoidal term Gk (i) is added:
signals (Fig. 6(b)).
Hk (i) = Hk + Gk (i) + k (i), (9)
To assuredly examine the existence of a signal component
that is caused by human breathing, we dive deeper into the Gk (i) = Ak cos(2fti + k ), (10)
measured signals to see whether there is a signal that has
where Ak , k , and f are the amplitude, phase, and frequency of
the same frequency as human breathing. Typically, an adult
the breathing induced periodic component in respective order
breathes at about 14 breaths per minute (bpm) at rest while
and ti is the time when the signal is sampled. For ease of
a newborn at 36 bpm [30], [31]. Thus the general breathing
presentation, this term of Gk (i) is also referred to breathing-
frequency of a static human is limited in 0.167 Hz to 0.667 Hz,
induced additive signal, or more concisely, breathing signal.
which corresponds to a breathing rate of 10 and 40 bpm,
Due to interference of human breathing, the sinusoidal term
respectively. Hence, we turn to investigating a potential signal
Gk (i) is generally larger than the noise term. In some ex-
from the received signals with a similar frequency.
treme cases with considerable environment noise, the breathing
To this end, we analyze the power spectral density (PSD) of
signal, however, could be drowned and thus potentially more
the filtered signal in the frequency domain. As shown in Fig. 7,
difficult to be identified. In this paper, we basically consider
we can clearly see a significant component that peaks at the
relatively stable environments, where the Gk (i) is larger than or
frequency of 0.233 Hz (about 14 bpm), which corresponds to
at least comparable to the noise term.
an adults typical breathing frequency (the true breathing rate
Intuitively, if the received signal is mean removed, it is
of this measurement is exactly 14 bpm). As a result, one can
possible to model the residual component with a sinusoidal
conclude the extracted signals as induced by human breathing
model. Further, human breathing can be detected by estimating
motion.
specific parameters (i.e., Ak , k , and f ) of the model. In other
words, given a deterministic model with unknown parameters,
B. Breathing Detection the problem of breathing detection is turned into a parameter
estimation problem, which can then be solved by optimization
To estimate breathing signals, we first pass the original techniques.
measurements through a filter to remove noises while keep the 3) Parameter Estimation: Specifically, the parameter es-
interested modes. Then we justify a sinusoidal model of the timation tasks are two-fold: 1) detect whether there exists
breathing signals and detect human breathing using a parameter a signal component that holds the interested frequency, and
estimation technique. 2) if yes, what is the specific frequency (and other parameters
1) Signal Filter: Before estimating breathing signal from like amplitude and period counts).
the mixed received signals, we first filter out the irrelevant Given the filtered signals Hk (i) sampled at time Ti , 1 i n,
components, such as electronic noise and other motions. We of which the mean amplitude is supposed to be removed, we
achieve this by applying a bandpass filter. aim to find a sinusoidal signal Gk (i) = Ak cos(2 f ti + k ) that
Since human breathing rates roughly fall into the range of minimizes the residual sum of squares (RSS)
10 bpm to 40 bpm, the filter should retain signals within
N 
 2
frequencies of 0.167 Hz and 0.667 Hz, while significantly  
RSS = Gk (i) Hk (i) (11)
attenuate the others. To be conservative, we set the minimum
i=1
frequency as fmin = 0.15 Hz and the maximum fmax = 0.70 Hz.
In addition, for a sequence of signals {Hk (i)}Ni=1 , the mean val- We tackle the problem using sinusoidal parameter estimation
ues Hk (i.e., the DC component) may bury the breathing com- algorithms that deal with a single sinusoidal signal of un-
ponent that usually only has low amplitudes of 1 dB or 2 dB. known frequency, phase, and amplitude. Particularly, we use the
The bandpass filter also removes the DC component in the orig- Nelder-Mead method [32], which is a common non-linear op-
inal signals, i.e., Hk = 0. The filtered signals then potentially timization technique for multidimensional unconstrained mini-
indicate the presence of human breathing. mization, i.e., minimizing an objective function with multiple
2336 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 11, NOVEMBER 2015

Fig. 8. Sinusoidal fitting on CFR amplitude at Subcarrier #10 a) User breaths evenly over the measuring period. (b) User breaths with different scale (yet similar
frequency) over the measuring period. (c) Noise signals without user breaths. (a) Human presence with f = 12 bpm. (b) human presence with f = 13 bpm.
(c) human absence (no breathing signals).

variables. Nelder-Mead algorithm is a simplex-based direct Ideally, human breathing should attach an additive signal
search that is effective and computationally compact and has with identical frequency (i.e., the breathing frequency) on each
been widely employed in parameter estimation and similar subcarrier, which is expected to be irrelevant to the signal
statistical problems. propagations. In practice, however, identical breathing motion
Fig. 8 illustrates the preliminary effects of the sinu- causes different extents of signal perturbation on different
soidal parameter estimation. As shown in Fig. 8(a), when subcarriers due to frequency-selective fading. Specifically, the
the person breathes evenly, a sinusoidal model can fit the same motion does not necessarily consistently increase or
breathing-induced signals with precise frequency and ampli- decrease the received signal power due to constructive and
tude. Fig. 8(b) portrays the case of a person breathing with destructive phaser superposition [8]. Hence breathing signals
inconsistent depths over time, which causes breathing sig- on some subcarriers would be more conspicuous and thus easier
nals with diverse amplitudes. As is shown, the sinusoidal to be captured while on others might be less obvious. Thus,
parameter estimation method still works excellently in such utilizing breathing interferences across multiple subcarriers
cases. Specifically, the frequency estimation remains accurate would improve both detection precision and robustness.
although the RSS of amplitude appears larger. We also test Taking Fig. 9 as an example, the rationale and necessity lie
the model estimation on non-breathing signals, i.e., signals in three folds:
measured when there is no human presence, and depict the 1) Breathing signals are well captured on most subcarriers,
results in Fig. 8(c). Compared with previous two figures, the yet with different amplitude responses. Fusing parameters
signals without breathing motion interference appear to be more on individual subcarriers would prospectively produce
random, lacking periodical patterns. When applying the param- more accurate estimation.
eter estimation algorithm, besides that the estimated frequency 2) Breathing motion may have no significant effects on a
is beyond the interested band of [fmin , fmax ], the deduced signals specific subcarrier, thus leading to miss detection using
make little sense since the amplitudes shrink to insignificance. that subcarrier only. Such miss detection can be avoided
Consequently, the estimation on non-breathing signals yields by incorporating results across multiple subcarriers.
meaningless results, which are remarkably discriminative from 3) In cases of human absence, there will be consistently no
those produced from breathing signals. significant periodic signals appearing on any subcarrier.
In a nutshell, we conclude that a sinusoidal model can fit the
breathing signals measured by CSI (in the form of individual To take advantages of multiple subcarriers, we first repeat
subcarrier) and the parameters can be precisely estimated by the the above sinusoidal parameter estimation for each individual
Nelder-Mead method. Thus one can successfully distinguish subcarrier. Then we obtain a group of breathing frequency
the presence and absence of a breathing person. In the follow- estimations fk , 1 k n, with the corresponding amplitude
ing, we will extend to the full frequency band, i.e., all available estimations Ak . Ideally, all fk should be the same since hu-
OFDM subcarriers, to make the detection more accurate and man breathing frequency is principally nondiscriminatory to all
robust. subcarriers. However, as discussed above, due to frequency-
selective fading, amplitude attenuation on each subcarrier
differs from each other, leading to inconsistent parameter es-
timation when fitting with a sinusoidal model. Hence we need
C. Embracing Frequency Diversity
to obtain reliable parameters from these primary estimations.
Modern modulations such as OFDM transmit information Diving into the parameter estimation on multiple subcarriers,
via multiple orthogonal subcarriers simultaneously to com- the majority of the frequency estimations are quite accurate
bat frequency-selective fading [33], giving rise to frequency and stable, and incorrect estimations occasionally appear, as
diversity for adaptive wireless communications [9], [33] and demonstrated by Fig. 10. Motivated by this observation, we
fine-grained indoor localization [24]. In this work, we also har- propose to sift out the biased incorrect frequency parame-
ness frequency diversity for more robust breathing parameter ters by conducting a one-dimension least median of squares
estimation. (LMS) outlier detection [34]. Mathematically, let f be the LMS
WU et al.: NON-INVASIVE DETECTION OF MOVING AND STATIONARY HUMAN WITH WiFi 2337

Fig. 9. Human breathing introduces diverse signal responses over different frequency subcarriers. (a) All subcarriers but a small portion (subcarrier #19#24)
depicts obvious breathing pattern, although the amplitude changes may be in the opposite direction (subcarrier #1#19 vs. subcarrier #25#30). (b) Estimated
signals are remarkably close to the breathing signals when using subcarrier #10. (c) Incorrect frequency with meaningless amplitude may be resulted in if using
subcarrier #24. (a) Breathing signals over all subcarriers. (b) breathing signals at subcarrier #5. (c) breathing signals at subcarrier #24.

Fig. 11. Experimental areas.

Fig. 10. Frequency estimation on a majority of subcarriers presents accurate

results while profaned estimations may appear on a small portion of subcarriers.
where I(fk ) is the outlier indicator we derived above. If f
Nevertheless, these biased estimations tend to be outliers (circled parts) and falls in a given frequency band of interests, [fmin , fmax ], then
could be sifted out by LMS estimator. a stationary human is detected; otherwise not.
Generally, LMS regression yields more reliable diagnosis
with more sample data. In practice, noticing that nowadays
estimate of f = [fk , k = 1, , n] and rk = fk f be the resid-
WiFi devices are often equipped with multiple antennas, we
uals, we determine whether fk is an outlier or not following a
incorporate multiple antennas to improve the reliability of
typical LMS regression as
frequency estimation. Specifically, we extend the sinusoidal

1 if |rk / | <= 2.5 parameter estimation to the CSI measurements on multiple
I(fk ) = (12) antennas, each of which results in n frequency estimations.
0 otherwise
Fusing all these estimations for the LMS regression, more
where accurate and reliable results can be gained than using a single
 antenna. Having said that, please note the proposed method
n benefits from, but does not rely on, multiple antennas.
qk rk2
= k=1
n
k=1 qk
 VI. E XPERIMENTS AND E VALUATION
1 if |rk /s0 | <= 2.5
qk = In this section, we first interpret the experiment methodology,
0 otherwise
followed by detailed performance evaluation.


5
s = 1.4826 1 +
0
medk rk2 .
n A. Implementation
The involved constant values are widely recognized factors in 1) Experimental Environments: We prototype DeMan with
the literature [34]. commodity WiFi devices and evaluate its performance in a
After diagnosing the biased and incorrect frequency estima- typical office building including a classroom and a laboratory
tions, we simply sift them out and then average the remained (Fig. 11). We employ a commercial TP-LINK TL-WR741N
ones as the ultimate estimation of breathing frequency: wireless router as the transmitter operating in IEEE 802.11n

n AP mode at 2.4 GHz. A LENOVO desktop running Ubuntu
1
f = n fk I(fk ) (13) 10.04 is used as a receiver, which is equipped with off-the-
k=1 I(fk ) k=1 shelf Intel 5300 NIC and a modified firmware [35]. During the
2338 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 11, NOVEMBER 2015

measurements, the receiver pings packets from the router and

records the CSI from each packet.
2) Experimental Methodology: We collect data from three
categories: 1) Moving data: There are one or more humans
walking around in the monitoring area. 2) Breathing data:
There is one human standing or sitting in the area of interests,
breathing naturally. 3) Human-free data: There is no human
presence and thus the environment is relatively static.
To collect moving data, we let a volunteer walk randomly
with a nature speed of around 1 m/s along a Hilbert-like
trajectory [36], which holds a space-filling property and thus
traverses the entire monitoring area. We adopt such trajectory
in purpose of demonstrating the sensitivity over the covered
space. Yet note that we did not necessarily follow a strict Hilbert Fig. 12. Detection accuracy using different features.
trajectory in the experiments. Instead, we let the volunteer take
a Hilbert-like but less complex path to traverse the space. CSI
data are continuously logged during the walking. For breathing
data, a volunteer stands or sits at a uniformed grid of locations
over the monitoring area, either on or off the LOS path between
the transmitter and receiver. For each grid location, we collect a
group of breathing data for around 2 minutes. For each link,
the monitoring area is defined as a rectangle area centered
along the TX-RX path with a width of 2 meter and a length
of the TX-RX distance. The volunteer keeps natural breathing
without other motions during the measurement. The ground
truth respirations are also manually counted and recorded. For
fair comparison, we collect a similar amount of data for each
category.
As shown in Fig. 11, we consider 5 different scenarios with Fig. 13. Impacts of packet quantity on detection accuracy of moving target.
different link conditions. Particularly, we consider both LOS
and NLOS propagation, diverse TX-RX distances ranging from
B. Performance
about 3 meters to 6 meters. For each scenario, we conduct
all the 3 types of measurements, each for three times. The 1) Performance of Detecting Moving Human: We first ex-
experiments are conducted in different days in a week, with amine and compare the performance of moving human detec-
about 8 hours of data or 540 k CSI records in total. tion using moving data and human-free data. Fig. 12 shows
For moving target detection, we first employ the well- the performance of moving detection, especially the gain from
known Support Vector Machine (SVM) classification to obtain exploitation of full CSI information, i.e., amplitude and phase.
a threshold line, based on a portion of measurements. The As is shown, DeMan achieves high TP and TN rates of
threshold line is one-time learned and, according to our ex- 94.07% and 98.55%, both outperforming the scheme with
periments, fits various scenarios. Yet different from previous merely CSI amplitude (by about 2% and 5% respectively).
work that conduct scenario-tailored calibration, we do not need To further demonstrate the benefits of using eigenvalues of
to calibrate the parameters for each different scenario over correlation matrices, we compare DeMan with a scheme that
different time. uses variances of amplitude and phase. As shown in Fig. 12,
3) Evaluation Metrics: We mainly use the following metrics the variance-based scheme attains poor performance with the
to evaluate the performance of DeMan. TP and TN rates of only 81.48% and 78.04% respectively.
We suspect that this is because variances, unlike eigenvalues,
True Positive Rate (TP): the fraction of cases where a are dependent on absolute signal powers and thus vary over
human (either moving or breathing) is correctly detected. different scenarios even when the human motions are similar,
True Negative Rate (TN): the ratio of cases where no making a pre-calibrated threshold line inapplicable for extended
human presence is correctly identified. cases.
We inspect the impacts of packet quantity by applying a
sliding window varying from 50 packets to 1000 packets. The
results are illustrated in Fig. 13, which demonstrates that TP
For each metric, we separately examine the performance of rate improves significantly with more packets (from 87.54%
moving human detection and stationary human detection, fol- with 50 packets to around 95% with more than 500 packets)
lowed with the overall evaluation. We also inspect the impacts while TN rate is almost unaffected at around 98%. The result is
of various factors, including packet quantity, link distance, LOS reasonable because in case of human-free, the channel measure-
and NLOS propagation, etc. ment mostly keeps stable and thus a short period of samples is
WU et al.: NON-INVASIVE DETECTION OF MOVING AND STATIONARY HUMAN WITH WiFi 2339

Fig. 14. Impacts of sample rate on detection accuracy of moving target. Fig. 16. Impacts of packet quantity on detection accuracy of static target.

Fig. 15. Stationary detection accuracy in diverse scenarios. Fig. 17. Impacts of sample rate on stationary target detection.

sufficient to capture the characteristics. In contrast, insufficient Fig. 16 illustrates the performance of stationary target detec-
samples fail to characterize the temporal variations of human tion with different amount of packets, given a fixed sample rate
motions since the influences from human movements do not of 50 Hz. As seen, DeMan achieves high detection accuracy,
uniformly distribute over time. Nevertheless, we observe that with TP and TN rates both higher than 95% with a window
satisfied accuracy of both TP and TN rates higher than 90% can of 1500 packets. The performance slightly degrades with more
be consistently achieved when using more than 200 packets. packets, since human breathing might vary and some motions
Fig. 14 shows the moving human detection performance could be involved during a long time window. Detection per-
under different sample rates. The results are drawn from data formance also drops with too fewer packets. Since human
measured within the same period of time window. As can be breathing motion is fairly slow, e.g., breathing once within a
seen, sample rate almost does not affect detection accuracy. period of a few seconds, observations within a very short period
This indicates that human movements continuously affects the make no sense of the breathing motion.
channel, and either frequent or sparse sampling can capture the The performance under different sampling rate is portrayed
variations well. in Fig. 17. Similar to moving human detection, sampling rate
2) Performance of Detecting Static Human: Breathing data has little impact on detection accuracy. Though a bit counter-
together with human-free data are used for evaluation of sta- intuitive, we suspect this is because measurements with sparse
tionary human detection. Fig. 15 demonstrates that DeMan sampling still depict the rhythmic patterns of breathing signals
achieves good detection performance in various scenarios as and thus the parameter estimation still works well. In other
depicted in Fig. 11, with TP and TN rates consistently higher words, as long as the signals periodicity is fully reserved,
than 90%. Integrating results from different cases, DeMan regardless of the specific sample rate, the performance of
provides a TP detection rate of 94.64% with a TN rate of breathing detection can be maintained.
94.49%. Even in case 4 as shown in Fig. 11 where the TX-RX We are particularly interested in the performance of
distance is more than 6 meters, the TP and TN rates still keep detection with stationary people presenting off the LOS path,
at 92% and 93%, respectively. Since the data are measured i.e., on the reflected paths. Fig. 18 depicts the detection rate
when the person stands or sits at different locations with diverse of people on and off the direct LOS path, with various packet
TX-RX distances, we conclude that DeMan extends the breath- numbers. As seen, DeMan achieves great performance in both
ing detection ability to cases of various poses, presenting conditions, with best TP and TN rates of above 95%. Such
locations and longer link lengths, compared with existing ap- encouraging results demonstrate the feasibility of DeMan in
proaches [21], [22]. various environments, without requiring users to stand directly
2340 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 11, NOVEMBER 2015

TABLE I
OVERALL P ERFORMANCE

Fig. 18. Detection performance in LOS and NLOS propagation.

Fig. 20. The upper figure displays accuracy of breathing rate estimation for
users on the LOS paths and the lower for users on the reflected paths.

and thus mixes up a significant portion (64.29%) of station-

ary cases with human-free scenarios. Although the baseline
Fig. 19. Results of motion indicator. The label confused indicates the
approach attains comparable TN rate, the corresponding false
measurements that are simultaneously inferred as motion and static. negative rate increases to a considerable level of 34.45%. These
results demonstrate that DeMan improves detection accuracy
and extends detection ability by exposing human breathing as
on or extremely close to the LOS path as demanded by previous
indicators for stationary humans. Note that the detection rate of
works [21], [22].
human-free scenarios is less than 99.86% as in Fig. 19 because
3) Putting it All Together: Finally, we examine the over-
that a small portion of human-free cases classified as static
all performance of DeMan in practical system with all three
could be falsely alarmed in the stationary target detection
categories of experimental data, integrating both moving hu-
module, which, however, is negligible.
man detection and stationary human detection with the motion
indicator process. We use the packet amount of 1500 for
evaluation. We implement a baseline approach for comparative VII. D ISCUSSIONS AND F UTURE W ORKS
study: a modified DeMan without the stationary target detection
A. Monitoring Breathing Rate
module, which can thus be treated as an improved version of
previous work [25] with phase information extension. Given that DeMan accurately detects stationary people by
We first present the results of primary motion indicator based sensing the breathing motion, we are also interested of how
on the variances of CSI as in Fig. 19. As seen, most of human- accurately DeMan could estimate the breathing rate. As shown
free cases are accurately categorized as static and only a very in Fig. 20, we test 35 groups of 1 minute measurements
small portion of 1.39% is labelled as in motion. In case of and observe that DeMan yields breathing rate of errors less
breathing human or moving human, however, a significant than 1 bpm for most testing cases. Concretely, the average
portion of around 35% will be falsely classified in the motion estimation error is 0.86 bpm for LOS scenarios and 0.97 bpm
indicator stage. Nevertheless, most of the falsely labelled cases for NLOS scenarios. These results demonstrate that DeMan
(25.32% out of 34.27% and 31.48% out of 34.81% for breathing achieves comparable accuracy of breathing rate estimation with
case and moving case respectively) will be doubly checked by previous contactless approaches and works effectively even
both moving and stationary detection. when the person presents on the reflected paths. In this sense,
Table I illustrates the overall detection results of DeMan. breathing rate monitoring is a side product of DeMan, which
Both moving people and stationary people can be accurately even outperforms previous dedicated works in case of longer
detected, with respective detection rate of 94.82% and 93.33%. link lengths and diverse user poses and presence locations
Integrating the results together, DeMan achieves TP and TN [20][22]. Nevertheless, currently we do not consider people
rates of 94.08% and 96.25%, respectively. As comparison, the diversity in our experiments, and it is of our interests to study
TP detection rate degrades to 65.55% for the baseline approach, how a persons age, gender, size, etc. affect breathing detection
which does not differentiate moving and stationary people and estimation.
WU et al.: NON-INVASIVE DETECTION OF MOVING AND STATIONARY HUMAN WITH WiFi 2341

B. Expanding Detection Coverage via Space Diversity performance for both moving and stationary people detection
Emerging generations of WiFi infrastructure are incorporat- in various environments. Requiring no prerequisite scenario-
ing an increasing number of antennas to boost capacity lever- tailored link calibration and being effective for persons on
aging space diversity [28]. Multiple antennas can be controlled and off the LOS paths, DeMan sheds lights on practical non-
digitally to adjust their beams towards a certain direction using invasive detection techniques.
beamforming techniques [37]. Such capability of narrowly
focusing transmission power on an intended direction can avoid R EFERENCES
the adverse impact of irrelevant multipath, and offer higher [1] J. Wilson and N. Patwari, See-through walls: Motion tracking us-
sensitivity on the blockage and reflection induced by the target ing variance-based radio tomography networks, IEEE Trans. Mobile
micro motions (breathing in our case) even from a NLOS path Comput., vol. 10, no. 5, pp. 612621, May 2011.
[2] C. Xu et al., SCPL: Indoor device-free multi-subject counting and locali-
[5]. In addition, since signals propagating through a NLOS path zation using radio signal strength, in Proc. ACM IPSN, 2013, pp. 7990.
often arrive at the receiver with a different angle-of-arrival, it [3] F. Adib, Z. Kabelac, D. Katabi, and R. Miller, 3D tracking via body radio
is possible to intentionally magnify the received signals from reflections, in Proc. USENIX NSDI, 2014, pp. 317329.
[4] Q. Pu, S. Gupta, S. Gollakota, and S. Patel, Whole-home gesture recogni-
directions of the NLOS paths to allow wider detection range, tion using wireless signals, in Proc. ACM MobiCom, 2013, pp. 2738.
and researchers have demonstrated the feasibility to distinguish [5] P. Melgarejo, X. Zhang, P. Ramanathan, and D. Chu, Leveraging direc-
arriving angles from multiple paths using commodity software tional antenna capabilities for fine-grained gesture recognition, in Proc.
ACM UbiComp, 2014, pp. 541551.
radios [38] and standard WiFi infrastructure [39]. [6] Y. Zhao, N. Patwari, J. M. Phillips, and S. Venkatasubramanian, Radio
tomographic imaging and tracking of stationary and moving people via
kernel distance, in Proc. ACM IPSN, 2013, pp. 229240.
C. Multiple Target Detection [7] J. Wilson and N. Patwari, Radio tomographic imaging with wireless net-
works, IEEE Trans. Mobile Comput., vol. 9, no. 5, pp. 621632, May 2010.
Current DeMan excels in detecting stationary humans in [8] J. Wilson and N. Patwari, A fade-level skew-laplace signal strength
case of individual person presence. In the future, we intend to model for device-free localization with wireless networks, IEEE Trans.
expand DeMan to multiple people scenarios. When there are Mobile Comput., vol. 11, no. 6, pp. 947958, Jun. 2012.
[9] D. Halperin, W. Hu, A. Sheth, and D. Wetherall, Predictable 802.11
multiple stationary persons in the monitoring area, each person packet delivery from wireless channel measurements, in Proc. ACM
is expected to contribute an additive signal with individual SIGCOMM, 2010, pp. 159170.
breathing frequency. Hence the resulted breathing signal will [10] Z. Yang, Z. Zhou, and Y. Liu, From RSSI to CSI: Indoor localization
via channel response, ACM Computing Surveys, vol. 46, no. 2, p. 25,
be a superposition of multiple sinusoidal signals. To extract Nov. 2013.
the breathing frequencies, we may then apply either a multiple [11] Y. Mostofi, Cooperative wireless-based obstacle/object mapping and see-
sinusoidal parameter estimation algorithm or successive inter- through capabilities in robotic networks, IEEE Trans. Mobile Comput.,
vol. 12, no. 5, pp. 817829, May 2013.
ference cancellation techniques [40]. Furthermore, we envision [12] A. Gonzalez-Ruiz, A. Ghaffarkhah, and Y. Mostofi, An integrated frame-
the impacts of chest motion on the phase information of CSI work for obstacle mapping with see-through capabilities using laser
as a potential proxy for more complicated cases where multiple and wireless channel measurements, IEEE Sensors J., vol. 14, no. 1,
pp. 2538, Jan. 2014.
persons could hold nearly identical breathing rates. [13] D. Huang, R. Nandakumar, and S. Gollakota, Feasibility and limits of
Wi-Fi imaging, in Proc. ACM SenSys, 2014, pp. 266279.
[14] S. Depatla, L. Buckland, and Y. Mostofi, X-ray vision with only WiFi
D. Extending to Through-Wall Detection power measurements using Rytov wave models, IEEE Trans. Veh.
The applicability of DeMan in case of human presence on Technol., vol. 64, no. 4, pp. 13761387, Apr. 2015.
[15] F. Adib and D. Katabi, See through walls with Wi-Fi! in Proc. ACM
the reflected paths encourages us to extend the detection ability SIGCOMM, 2013, pp. 7586.
to NLOS scenarios, i.e., the LOS propagation is prohibited [16] B. Kellogg, V. Talla, and S. Gollakota, Bringing gesture recognition to
between the TX and RX. Specifically, we are interested in all devices, in Proc. USENIX NSDI, 2014, pp. 303316.
[17] Y. Wang et al., E-eyes: In-home device-free activity identification using
exploring the possibility of through-wall detection of human fine-grained WiFi signatures, in Proc. ACM MobiCom, 2014, pp. 617628.
presence, yet without involving a dense network or prior link [18] F. Q. AL-Khalidi, R. Saatchi, D. Burke, H. Elphick, and S. Tan, Respi-
calibration as [1], [7]. The potentiality of detecting people in ration rate monitoring methods: A Review, Pediatric Pulmonol., vol. 46,
no. 6, pp. 523529, Jun. 2011.
case of NLOS propagation is critical and helpful for various ap- [19] C. Li, J. Ling, J. Li, and J. Lin, Accurate doppler radar non-contact vital
plications like disaster relief where survivors are often trapped sign detection using the RELAX algorithm, IEEE Trans. Instrum. Meas.,
behind the wall or buried underground. vol. 59, no. 3, pp. 687695, Mar. 2010.
[20] N. Patwari, J. Wilson, S. Ananthanarayanan, S. Kasera, and D. Westenskow,
Monitoring breathing via signal strength in wireless networks, IEEE
VIII. C ONCLUSION Trans. Mobile Comput., vol. 13, no. 8, pp. 17741786, Aug. 2014.
[21] O. J. Kaltiokallio, H. Yigitler, R. Jntti, and N. Patwari, Non-invasive
In this work, we present the design and implementation of respiration rate monitoring using a single COTS TX-RX pair, in Proc.
DeMan, a unified framework for simultaneous detection of ACM IPSN, 2014, pp. 5970.
[22] X. Liu, J. Cao, S. Tang, and J. Wen, Wi-sleep: Contactless sleep moni-
moving and stationary people with COTS WiFi devices. We ex- toring via WiFi signals, in Proc. IEEE RTSS, 2014, pp. 346355.
ploit both amplitude and phase information of CSI for moving [23] K. Wu, J. Xiao, Y. Yi, M. Gao, and L. M. Ni, FILA: Fine-grained indoor
target detection. To detect stationary people, we leverage the localization, in Proc. IEEE INFOCOM, 2012, pp. 22102218.
[24] S. Sen, B. Radunovic, R. R. Choudhury, and T. Minka, You are facing
breathing-induced periodic patterns on wireless signals using a the Mona Lisa: Spot localization using PHY layer information, in Proc.
sinusoidal model. To improve detection accuracy and robust- ACM MobiSys, 2012, pp. 183196.
ness, we further harness the frequency diversity across multiple [25] J. Xiao, K. Wu, Y. Yi, L. Wang, and L. Ni, FIMD: Fine-grained device-
free motion detection, in Proc. IEEE ICPADS, 2012, pp. 229235.
OFDM subcarriers. We prototype DeMan in typical buildings. [26] W. Xi et al., Electronic frog eye: Counting crowd using WiFi, in Proc.
Experimental results demonstrate that DeMan achieves great IEEE INFOCOM, 2014, pp. 361369.
2342 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 11, NOVEMBER 2015

[27] M. Gast, 802.11 Wireless Networks: The Definitive Guide. Sebastopol, Zimu Zhou (S13) received the B.E. degree in 2011
CA, USA: OReilly Media, Inc., 2005. from the Department of Electronic Engineering,
[28] D. Halperin, W. Hu, A. Sheth, and D. Wetherall, 802.11 with multiple an- Tsinghua University, Beijing, China. He is currently
tennas for dummies, ACM SIGCOMM Comput. Commun. Rev., vol. 40, pursuing the Ph.D. degree in the Department of
no. 1, pp. 1925, Jan. 2010. Computer Science and Engineering, Hong Kong
[29] L. Davies and U. Gather, The identification of multiple outliers, J. Amer. University of Science and Technology, Kowloon,
Stat. Assoc., vol. 88, no. 423, pp. 782792, Sep. 1993. Hong Kong. His main research interests include
[30] P. Sebel, M. Stoddart, R. Waldhorn, C. Waldman, and P. Whitfield, Respi- wireless networks and mobile computing. He is a
ration, the Breath of Life. New York, NY, USA: Torstar Books, 1985. student member of the ACM.
[31] J. F. Murray, The Normal Lung: The Basis for Diagnosis and Treatment
of Pulmonary Disease. Philadelphia, PA, USA: Saunders, 1986.
[32] J. C. Lagarias, J. A. Reeds, M. H. Wright, and P. E. Wright, Convergence
properties of the Nelder-Mead simplex method in low dimensions, SIAM
J. Optim., vol. 9, no. 1, pp. 112147, 1998. Xuefeng Liu received the M.S. and Ph.D. degrees
[33] A. Bhartia, Y.-C. Chen, S. Rallapalli, and L. Qiu, Harnessing frequency from Beijing Institute of Technology, Beijing, China,
diversity in Wi-Fi networks, in Proc. ACM MobiCom, 2011, pp. 253264. and University of Bristol, U.K., in 2003 and 2008,
[34] P. J. Rousseeuw and A. M. Leroy, Robust Regression and Outlier Detec- respectively. He is currently a Research Fellow with
tion. Hoboken, NJ, USA: Wiley, 2005. the Department of Computing, Hong Kong Poly-
[35] D. Halperin, W. Hu, A. Sheth, and D. Wetherall, Tool release: Gath- technic University, Kowloon, Hong Kong. His re-
ering 802.11n traces with channel state information, ACM SIGCOMM search interests include wireless sensor networks,
Comput. Commun. Rev., vol. 41, no. 1, pp. 53, Jan. 2011. distributed computing and in-network processing.
[36] Y. Ling, S. Alexander, and R. Lau, On quantification of anchor place- He has served as a reviewer for several international
ment, in Proc. IEEE INFOCOM, 2012, pp. 21922200. journals/conference proceedings.
[37] X. Zhou et al., Mirror mirror on the ceiling: Flexible wireless links for
data centers, in Proc. ACM SIGCOMM, 2012, pp. 443454.
[38] J. Xiong and K. Jamieson, ArrayTrack: A fine-grained indoor location
system, in Proc. USENIX NSDI, 2013, pp. 7184. Yunhao Liu received the B.S. degree in automation
[39] J. Gjengset, J. Xiong, G. McPhillips, and K. Jamieson, Phaser: Enabling from Tsinghua University, Beijing, China, in 1995,
phased array signal processing on commodity WiFi access points, in and the M.S. and Ph.D. degrees in computer science
Proc. ACM MobiCom, 2014, pp. 153164. and engineering from Michigan State University,
[40] S. Gollakota and D. Katabi, Zigzag decoding: Combating hidden termi- in 2003 and 2004, respectively. He is currently a
nals in wireless networks, in Proc. ACM SIGCOMM, 2008, pp. 159170. ChangJiang Professor at Tsinghua University. His
research interests include wireless sensor network,
peer-to-peer computing, and pervasive computing.
Chenshu Wu (S12) received the B.S. degree
from the School of Software, Tsinghua University,
Beijing, China, in 2010. He is now pursuing the
Ph.D. degree in the Department of Computer Science
and Technology, Tsinghua University. His research
interests include wireless ad-hoc/sensor networks
Jiannong Cao (M93SM05F15) received the
and mobile computing. He is a student member of
B.Sc. degree in computer science from Nanjing Uni-
the ACM.
versity, Nanjing, China, in 1982, and the M.Sc. and
Ph.D. degrees in computer science from Washington
State University, Pullman, WA, USA, in 1986 and
1990, respectively. He is currently the Head and
Chair Professor of the Department of Computing,
Hong Kong Polytechnic University, Hong Kong. Be-
Zheng Yang (S07M11) received the B.E. degree fore joining Hong Kong Polytechnic University, he
in computer science from Tsinghua University, Bei-
was on the Faculty of Computer Science at James
jing, China, in 2006 and the Ph.D. degree in com-
Cook University, the University of Adelaide in Aus-
puter science from Hong Kong University of Science tralia, and the City University of Hong Kong. His research interests include par-
and Technology, Kowloon, Hong Kong, in 2010. He allel and distributed computing, networking, mobile and wireless computing,
is currently an Assistant Professor at Tsinghua Uni-
fault tolerance, and distributed software architecture. He has published more
versity. His main research interests include wireless
than 200 technical papers in the above areas. His recent research has focused on
ad-hoc/sensor networks and mobile computing. He mobile and pervasive computing and mobile cloud computing. He is a senior
is a member of the ACM. He was awarded the 2011
member of the China Computer Federation, a Fellow of the IEEE Computer
National Nature Science Award (second class).
Society, and a member of the ACM.