A WiFi-based Home Security System

Abstract

Typical home security systems monitor homes for intrusions by installing contact sensors on doors and windows and motion sensors inside the house. Unfortunately, due to the
high deployment and operational costs of today’s home security systems, only a small fraction of homes have security systems in- stalled (e.g., only 17% in the US and 15% in China). In this paper, we propose a WiFi based Home Security system (WiHS) that
uses commodity WiFi devices, which most modern households already have, to perform the three primary tasks of typical home security systems: 1) detect when a door/window is opened/closed, 2) identify which door/window has been opened/closed, and 3)
detect movements inside the house. The design of WiHS is based on our intuitive and theoretical understanding of the impacts of the movements of doors and windows on WiFi signals, which we will develop and present in this paper. We extensively evaluated
WiHS using commodity WiFi devices in 3 different houses. WiHS detected intrusions with over 95% accuracy and identified the exact door/window that moved with just 4.5% average error.

Full text

A WiFi-based Home Security System
Shaohu Zhang, Raghav H. Venkatnarayan, Muhammad Shahzad
North Carolina State University, USA. {szhang42, rhampap, mshahza}@ncsu.edu
Abstract—Typical home security systems monitor homes for
intrusions by installing contact sensors on doors and windows
and motion sensors inside the house. Unfortunately, due to the
high deployment and operational costs of today’s home security
systems, only a small fraction of homes have security systems in-
stalled (e.g., only 17% in the US and 15% in China). In this paper,
we propose a WiFi based Home Security system (WiHS) that
uses commodity WiFi devices, which most modern households
already have, to perform the three primary tasks of typical home
security systems: 1) detect when a door/window is opened/closed,
2) identify which door/window has been opened/closed, and 3)
detect movements inside the house. The design of WiHS is based
on our intuitive and theoretical understanding of the impacts of
the movements of doors and windows on WiFi signals, which we
will develop and present in this paper. We extensively evaluated
WiHS using commodity WiFi devices in 3 different houses. WiHS
detected intrusions with over 95% accuracy and identified the
exact door/window that moved with just 4.5% average error.
I. INTRODUCTION
Home security systems provide good deterrence against
intrusions and burglaries [1], [2]. A typical home security
system installs contact sensors on doors and windows and
motion sensors inside the house, and raises an alarm if a
door/window is opened or any motion is detected. Unfortu-
nately, the deployment cost of today’s home security systems
is high. For example, in the US, the cost of a typical door,
window, and motion sensor ranges from $30 to $80 [3], [4].
A home with front and back entrance doors, 10 windows, and
2 motion sensors incur about $500 for just the senors; security
monitoring hub and the labor cost are in addition. Due to such
high costs, only a small percentage of homes have security
systems [5] (e.g.,<17% in US [6] and <15% in China [7]).
While a large percentage of homes do not have security
systems, most modern homes have WiFi. If one could replicate
the monitoring functions of a conventional security system
using existing WiFi devices, the resulting WiFi based security
system will come at negligible/no monetary cost. Researchers
have already shown that different human movements impact
WiFi signals differently, and have leveraged this observation to
develop human sensing systems [8]–[12]. We further observed
that the movements of doors and windows situated at different
locations in a house also impact WiFi signals differently. Thus,
by measuring changes in WiFi signals, we should be able to
detect whether and which door/window is being opened/closed
and whether there is a movement inside the house.
Problem Statement: Our objective is to develop a WiFi based
home security system that can perform the three primary
monitoring tasks of the conventional security systems: 1)
detect when a door/window is opened/closed, 2) identify which
door/window has been opened/closed, and 3) detect whenever
there is a movement inside the house.
Proposed Solution: In this paper, we propose WiHS, a WiFi
based Home Security system that can accurately perform the
three monitoring tasks using commodity WiFi devices. We
have designed WiHS for two typical scenarios: 1) away-
mode, where all occupants have gone outside (such as to
their offices/schools), and 2) stay-mode, where one or more
occupants are inside the house (such as when going to bed
at night). When arming WiHS, an occupant selects which
scenario he/she is arming it for.
Fig. 1 shows a block diagram that illustrates how WiHS uses
the channel state information (CSI) reported by commodity
WiFi devices to perform the three monitoring tasks. The
denoising block continuously captures the CSI measurements
from WiFi network interface cards (NIC), removes noise from
them, and outputs a time series of denoised CSI values,
which we will refer to as denoised-stream. As noise removal
from CSI is a well studied topic, we simply adopted the
principal component analysis based noise removal technique
proposed in [13], [14], and will thus not discuss it further.
The segmentation block takes the denoised-stream as input and
continuously analyzes it to detect any movement. As soon as it
detects a movement, it first determines whether the movement
happened inside the house or outside, and if it is inside, it
then determines whether the detected movement was due to a
moving door/window or human. If the movement was due to
a door/window, it raises an alarm and outputs the segmented
portion of the denoised-stream containing the movement. If the
movement was due to a human, it raises an alarm only if the
system is armed for away-mode. The D/W–identification block
takes the segmented denoised-stream as input and evaluates it
against the classification model of each door and window in
the house to determine which door/window moved. Finally, the
O/C–identification block takes the same segmented denoised-
stream along with the decision of the D/W–identification block
as inputs, evaluates it against the classification model of open
and close events of the identified door/window to determine
whether the door/window opened or closed.
O/C Training Block
Database
CSI
streams of
training
samples
oising
D/W Training Block
Denoised
streams
of training
samples
O/C trained models
D/W trained models
Training
Runtime
Fig. 1. Block Diagram of WiHS
To generate classification models that the D/W– and O/C–
identification blocks use, for each door/window, WiHS re-
quires an occupant to provide about 30 training samples
each of opening and of closing that door/window. WiHS

passes these samples through the denoising block and provides
the denoised-streams as input to the D/W–training and the
O/C–training blocks. From these denoised-streams, the D/W–
training block extracts features appropriate for distinguishing
between doors and windows, trains classifiers using the ex-
tracted features, and stores the trained classification models in
a database from where the D/W–identification block retrieves
them at runtime. The O/C–training block trains classifiers in
a similar way for the O/C–identification to use at runtime.
Positioning the Proposed Solution: Compared to conven-
tional security systems, which provide deterministic outputs,
WiHS is probabilistic and can miss or misclassify some
events. We emphasize that WiHS is not intended to replace
conventional security systems. It is intended to provide a
no/low-cost solution using existing WiFi equipment for homes
that otherwise do not have a security system at all.
Key Contributions: In this paper, we make five key con-
tributions. 1) We present WiHS, a WiFi based home se-
curity system that performs the three monitoring tasks of
conventional security systems using commodity WiFi devices.
2) We characterize the impact of the movements of doors
and windows on WiFi signals and explain our observations
both intuitively and theoretically. 3) We present a method
to distinguish between the movements of humans and of
doors/windows inside any house. 4) We present our imple-
mentation and extensive evaluations of WiHS in three different
houses. 5) We will open source the large data set of WiFi traces
that we have collected and used to evaluate WiHS.
II. MOV EM EN T SEGMENTATION
Let NT x and NRx represent the number of T x and Rx
antennas, respectively, on any given WiFi device. Let S
represent the number of OFDM subcarriers between each T x-
Rx pair. Each CSI measurement comprises S×NTx ×NRx
channel frequency response (CFR) values, one for each subcar-
rier between each T x-Rx pair. As WiFi NICs generate CSI
measurements repeatedly, we obtain S×NTx ×NRx time-
series of CFR values. Onward, we will call each time-series
of CFR values a CSI-stream.
1) Segmentation: To detect the start and end of a move-
ment, on any given denoised-stream, the segmentation block
(shown in Fig. 1) slides a window of size Wwith a step size
of Sand calculates the variance of the values covered by the
window in each step. We call the resulting stream of values a
variance-stream. Fig. 2(a) shows a denoised-stream resulting
from a door open followed by a close followed by another open
and close. Fig. 2(b) shows the corresponding variance-stream.
(a) Denoised CSI-stream
Amplitude
(b) Variance-stream
Variance
Open Close Open
Close
Fig. 2. Denoised and vairance-streams
from door open and close events
We observe from this fig-
ure that the values in
the variance-stream in the
presence of a door move-
ment are much higher
compared to in the ab-
sence. We made similar
observations for window
and human movements.
To detect the start and end of a movement, the segmentation
block first identifies all the local maxima in the variance-
stream by comparing each value with a value before and after
it. Next, it determines whether each local maximum represents
a prominent peak or if it just appeared due to noise. For this,
it compares the value of each maximum with a threshold T
(we will describe shortly how WiHS sets the value of T). This
simple approach works because the variance in the presence
of a movement is much larger compared to in the absence.
The segmentation block discards all the local maxima whose
values are less than T. We refer to the remaining local maxima
as peaks. The markers in Fig. 2(b) show the peaks detected
by the segmentation block using this approach. Next, the
segmentation block randomly chooses a peak and identifies
all those peaks in its vicinity that are separated in time by
less than a second from their respective adjacent peaks, and
obtains a set of closely spaced peaks. We empirically observed
that in the presence of a human movement or a door/window
movement, the adjacent peaks are never separated by more
than a second. The segmentation block uses the first point
in the denoised-stream covered by the sliding window that
generated the first peak in the obtained set of peaks as the
start point of the movement. Similarly, it uses that point in
the denoised-stream as the end point of the movement that
comes Wpoints after the last point in the denoised-stream
covered by the sliding window that generated the last peak in
the obtained set of peaks. The markers in Fig. 2(a) show the
start and end points detected using this method.
To set the value of T, at the time of initial setup,
WiHS requires the homeowner to perform a one-time calibra-
tion operation. In this operation, it requires the homeowner
to first keep the environment inside the house static for
20 seconds by ensuring that no humans or doors/windows
inside the house move, and then randomly walk around
in the house for another 20 seconds. After this, WiHS
calculates mean µSof the variance-stream correspond-
ing to the 20 seconds of no movement and mean µW
0.05 0.1 0.15 0.2 0.25 0.3
0.5
0.6
0.7
0.8
0.9
1
CDF
Static
Non-walk
Walk outside
Fig. 3. %age of values ≤Tvs. α
of the variance-stream corre-
sponding to the 20 seconds of
walk and sets T= (1 −α)
×µS+α×µW. Next, we de-
termine the appropriate value
of the weighting factor α.
Fig. 3 plots the percentage of values that are below Tin
the variance-streams in all our traces when the environment
was static, where we varied αfrom 0.05 to 0.3. We observe
from this figure that when α≥0.1, 100% of the values in the
variance streams are < T when the environment was static.
Fig. 3 also plots the percentage of values that are below T
when one or more persons walked outside (but close to the
outer walls of) the house. We observe that when α≥0.15,
100% of the values in the variance streams are < T when
the people moved outside the house. Fig. 3 further plots the
percentage of values that are below Twhen one or more
persons performed non-walking movements (such as sitting
down, eating, etc.) inside the house. We observe that when

α≥0.25, 98.34% of the values in the variance streams
are < T when people performed non-walking movements
inside the house. We further observed from our traces that
when α= 0.25, over 85.65% of the local maxima in the
variance-stream were greater than Twhen a door/window
moved or a human walked inside the house. Consequently,
we set α= 0.25. To visually demonstrate these observations,
consider Figs. 4(a) and 4(b), where we plot the denoised-
stream and corresponding variance stream, respectively, when
Amplitude
(a) Denoised CSI-stream
Variance
(b) Variance-stream
Walk
outside Open Close
Walk
inside
Fig. 4. Denoised and variance-streams
resulting from various activities
a user performed four activ-
ities: 1) briefly walked out-
side the house, 2) opened
the door, 3) entered and
closed the door, and 4)
walked inside the house. We
can clearly see from this
figure that walking outside
the house results in significantly smaller (almost invisible)
peaks in the variance-stream compared to walking inside the
house or opening/closing a door. Thus, using α= 0.25,
the segmentation block is able to ignore movements that
happen outside the house and any non-walking movements
that happen inside, and detects the start and end of only
those movements that are comprised of human walk or the
door/window movement.
2) Distinguishing Door/Window vs. Walking: The method
described above already enables the segmentation block to de-
tect start/end times of movements that are due to door/window
or walk, and enables it to discard any non-door/window/walk
movements inside the house or any movements outside. Next,
we describe how it determines whether a segmented movement
is due to a person walking (or significantly changing position)
or due to a door/window.
WifiU [15] and WiStep [16] demonstrated that human
movements introduce frequencies up to 20 Hz in CSI at 2.4
GHz. We will show (intuitively, theoretically, and experimen-
tally) in Secs. III-A and III-B that the movements of doors
and windows introduce frequencies in the range of 3 to 11
Hz in the denoised-streams. These observations are visually
demonstrated in Figs. 5 and 6 as well, where we can see that
the door/window movements do not introduce any significant
frequencies beyond 11 Hz in the denoised-stream, whereas,
human walk introduces frequencies beyond 11 Hz as well. The
segmentation block leverages these observations to distinguish
between door/window movements and walking movement.
2 5 10 15 20 25 30
Frequency (Hz)
0
0.2
0.4
0.6
0.8
1
Magnitude
Fig. 5. Freqs. introduced by door
2 5 10 15 20 25 30
Frequency (Hz)
0
0.2
0.4
0.6
0.8
1
Magnitude
Fig. 6. Freqs. introduced by walk
Given any segmented denoised-stream, the segmentation
block first applies fourier transform on it. Next, it performs
max-min normalization on the magnitudes of the frequencies
in the range of 3 to 20 Hz to bring the magnitude of each
frequency in the range of 0 to 1. After this, it calculates the
sum of the magnitudes of the frequencies in the range of
3 to 11 Hz and divides it with the number of frequencies
in this range. We represent the resulting value with PD/W.
PD/W, in essence, is the power spectral density over the
range of 3 to 11 Hz. Similarly, the segmentation block also
calculates PWalk, which is the sum of the magnitudes of the
frequencies in the range of 11 to 20 Hz divided by the number
of frequencies in this range. If the movement in the given
segmented denoised-stream is a door/window movement, then
as per our observations in the last paragraph, PD/W must be
much greater than PWalk compared to if the the movement is
a walking movement. Fig. 7 shows the box plots of the ratio
D/W Walking
0
50
100
150
Fig. 7. PD/W/PWalk for
D/W and walk
PD/W/PWalk for the samples in our traces.
We can see that this ratio is much larger
for door/window movements compared
to human walk. To determine whether
the movement in the given segmented
denoised-stream is a human movement
or a door/window movement, the seg-
mentation block compares the ratio PD/W/PWalk with a thresh-
old R. If the ratio is greater than R, the segmentation block
identifies the movement in the given segmented denoised-
stream as a door/window movement and raises an alarm. If
the ratio is less than R, it identifies the movement as a human
walk and raises an alarm if WiHS has been armed in the away-
mode. We have empirically determined the value of Rto be
20, as we will show later in the evaluation section.
III. DOOR/WIN DOW ID EN TI FIC ATIO N
A. Intuition
As different doors and windows are situated at different
angles from the WiFi receiver in any given house, the varia-
tions that their movements introduce in the CSI measurements
are different. This intuition follows from the observation that
Virmani et al. demonstrated in [10] that when a human
performs any given gesture while facing in different directions,
the variations introduced in the CSI measurements by the
same gesture are different. As an example, Fig. 8(a) plots
two denoised-streams from two samples of fully opening
door 1 in House # 1 (floor plans will be described later
in Fig. 12) and Fig. 8(b) plots two denoised-streams from
two samples of fully opening door 2. Similarly, Figs. 8(c)
and 8(d) plot two denoised-streams from two samples each
of fully opening window 1 and window 2, respectively, in
House # 1. We observe from Figs. 8(a) and 8(b) that the
denoised-streams of the two doors are different from each
other due to the difference in their relative positions from the
WiFi receiver. We make similar observations from Figs. 8(c)
and 8(d) for the two windows. We further observe that for
any given door/window, the denoised-streams from the two
samples of that door/window are fairly similar. Consequently,
(a) Door 1 (b) Door 2 (c) Window 1 (d) Window 2
Fig. 8. Denoised-streams from two samples of opening

as the movement of any given door/window generates similar
variations in the denoised-stream across multiple samples
from the same door/window, and as these variations are very
different for different doors/windows, by first learning the
patterns of change in the denoised-stream introduced by the
movement of any given door/window using some training
samples, we can use an appropriate machine learning classifier
to identify that door/window when it moves at runtime.
B. Feature Extraction
1) Quantifying the Impact of Door/Window Movements on
CFR: A transmitted signal arrives at the receiver from multiple
paths. Let ak(f, t)be the complex-valued representation of the
initial phase and the attenuation of the kth path at time tfor a
signal with carrier frequency f. Let dk(t)represent the length
of the kth path at time t. We can represent this aggregate
CFR as the sum of a dynamic and a static component. The
dynamic component changes as a door/window moves and
is the sum of the CFRs of all those paths that arrive at the
receiver after reflecting from the moving door/window in the
environment. Let Pdrepresent the set of all those paths that
reflect from the moving door/window and arrive at the receiver.
The static component is not affected by the movement of
any door/window and is the sum of the CFRs of all those
paths that arrive at the receiver without reflecting from any
moving objects. Let Hs(f)represent the static component of
the aggregate CFR. Let vkrepresent the rate at which the
length of the kth path changes. We call vkthe speed of the kth
path. Thus, dk(t) = dk(0) + vkt. The aggregate CFR power
of the signal arriving at the receiver is given by the following
well-known equation [14] [11]:
|H(f, t)|2=
k∈Pd
2|Hs(f)ak(f, t)|cos 2πvkt
λ+2πdk(0)
λ+φsk
+
k,l∈Pd;k�=l
2|ak(f, t)al(f , t)|cos 2π(vk−vl)t
λ+2π(dk(0)−dl(0))
λ+φkl
+
k∈Pd
|ak(f, t)|2+|Hs(f)|2(1)
where 2πdk(0)/λ+φsk and 2π(dk(0) −dl(0)) /λ+φkl are
constants representing initial phase offsets. We observe from
Eq. (1) that the only sources of periodic variations in the total
CFR power are the two terms 2πvkt/λand 2π(vk−vl)t/λ,
and the frequencies in these periodic variations are determined
by vkand vl. Thus, the frequencies of the sinusoids in the total
CFR power are functions of the speeds of path length changes.
To illustrate this, consider a typical 0.9 meters (∼3 ft) wide
door. From our data sets (which we will describe later in Sec.
V), we observed that typically it takes people 2 to 6 seconds
to open a door by 90◦. Thus, the speed vkof the signal path
from the door to the WiFi receiver generally lies in the range
of about 2×2×π/4×0.9
6to 2×2×π/4×0.9
2∼0.47 to 1.41 m/s.
As per the first cosine term in Eq. (1), the frequencies that
should appear in the CFR power as the door moves should be
vk/λ,i.e.,≈3to 11 Hz (using λ≈0.125 m for the 2.4 GHz
WiFi band). Figs. 9 and 10 show spectrograms obtained from
the denoised-streams resulting from the opening and closing of
a door and of a window, respectively. To obtain a spectrogram
from any given denoised-stream, we apply short time fourier
transform (STFT) on it. STFT slides a window over the
denoised-stream, where at each sliding step, it applies fast
fourier transform (FFT) on the values covered by the window
in that step. As our sampling rate is 300 samples/sec, we chose
the window size of 300 samples and step size of 15 samples,
which resulted in a good frequency resolution of 1 Hz and time
resolution of 50 ms. An FFT at any given sliding window
step results in a vector of magnitudes of all frequencies in
the portion of the denoised-stream covered by the window
in that step. We will refer to this vector as FFT vector. As
door movements give rise to frequencies less than 11 Hz,
we only use FFT values of the first 11 Hz. A spectrogram
of any given denoised-stream is essentially a concatenation
of all FFT vectors obtained from sliding the window over
that denoised-stream. We indeed observe from Figs. 9 and 10
that the frequencies that appear with high magnitude in the
spectrogram lie under approximately 11 Hz. We do see some
frequencies >11 Hz but they have low magnitude, and appear
due to faster moving body parts of the person open/closing the
door/window. We also see frequencies <3Hz, which appear
permanently due to the DC or slowly varying components in
the denoised-stream. These spectrograms empirically validate
the impact of door movement on denoised-stream that we
theoretically predicted above using Eq. (1).
0 5 10 15 20 25
Time (s)
0
5
10
15
20
Frequency (Hz)
Fig. 9. Spectrogram resulting from
the opening and closing of a door
0 5 10 15 20 25
Time (s)
0
5
10
15
20
Frequency (Hz)
Fig. 10. Spectrogram resulting from
the opening and closing of a window
Takeaway: As different doors are situated at different dis-
tances and angles from the receiver, different doors have
different relative speeds seen by the receiver as they
move, due to which, they introduce different frequencies in
denoised-streams, which makes denoised-streams from differ-
ent doors/windows different. Thus, for the purpose of distin-
guishing across different doors and windows, the best features
to use are the magnitudes of different frequencies that appear
in the denoised-streams as any door/window moves.
2) Extracting the Features: From Fig. 9, we observe that
both door open and close give rise to the same set of frequen-
cies in the CFR power, just that the FFT vectors appear in the
reverse order due to the opposite directions of motion. Thus, if
we take the average of the magnitudes of any given frequency
during a door open/close event across all the FFT vectors, that
average should be the same whether the door was opened or
closed. The black lines in Fig. 11 show the average magnitude
of each frequency during the open and close events shown in
Fig. 9. The gray lines show the average magnitudes during an
open and a close event of another door. The figure shows that
for any given door, regardless of whether the door opens or
closes, the average magnitude of each frequency is fairly equal.
However, across different doors, the average magnitudes of

1 5 10 15 20
Frequency (Hz)
50
60
70
80
90
100
Avg. Magnitude
Door 1
Door 2
Fig. 11. The mean and median of
frequency magnitudes for opening
and closing of two doors
any given frequency differ.
This is inline with our take-
away. We made similar obser-
vations when using median val-
ues instead of the mean values.
Thus, the D/W–training block
uses the mean and median of
each frequency in the range of
3∼11 Hz as features. More specifically, for each door/window
open/close event, it generates the spectrogram using the values
in the denoised-stream from the start till the end of the event.
As a spectrogram has multiple rows (one per frequency), it
calculates a mean and a median for each row starting from
row # 3 till row # 11, and uses these values as features.
Note that as the frequencies that appear when a door
opens/closes depend on the speed with which the door opens,
we must handle the changes in speed. To make the duration
of all door/window movements consistent, the D/W–training
block virtually expands/contracts the given sample. Let T
represent the duration of the sample. If T>1sec, the D/W–
training block virtually contracts the sample by T−1seconds
to normalize its duration to one second, and vice versa. Ex-
panding/contracting a sample implies that the door was moved
at a slower/faster rate, which in turn implies that the frequen-
cies in the aggregate CFR power must be decreased/increased
depending on the extent of expansion/contraction of the given
sample. To achieve this, the D/W–training block re-scales the
y-axis of the spectrogram by simply multiplying it with T. To
see why, consider a door that moves by distance din time T.
Thus the speed vof the change in path length is 2d
T, and the
frequency Fthat appears in the CFR power is F=v
λ=2d
Tλ.
Rearranging this, we get d=TλF
2. If the same door movement
was to happen in 1 sec, the distance moved by the door is
same d. Let us represent the frequency resulting from this
door movement by f. Through the same derivation as above,
we get d=λf
2for time 1sec and frequency f. Equating the
two equations, we get f=F× T .
C. Classifier Training
The D/W–training block generates an independent single
class classification model for each door/window. The mo-
tivation behind using single class classifiers is that due to
the differences in the physical positions of the doors and
windows, different doors and windows give rise to different
sets of frequencies in the range of 3 to 11 Hz that are more
prominent. If we use frequencies that are more tailored for
each door, naturally, we get higher accuracy. To select the
appropriate frequencies for any given door/window, the D/W–
training block takes the mean and median values for each
frequency, trains a classier (classifier training will be discussed
shortly) using the two-valued feature vector from each training
sample of that door/window and calculates the true positive
rate of the resulting classifier through 10-fold cross validation.
Next, the D/W–training block discards all those frequencies
for this door/window that attain less than 80% true positive
rate. After selecting all frequencies for the given door/window,
the D/W–training block uses their medians and means as
features to train the classifier for this door/window. We call
the set of mean and median values of the selected frequencies
as feature vector. In our implementation, we chose Support
Vector Distribution Estimation (SVDE) with the Radial Basis
Function (RBF) kernel as our single class classification model.
To generate an SVDE classification model for any given
door/window, the D/W–training block first normalizes the val-
ues of each feature in the feature vectors of that door/window
from all training samples of that door/window to bring them
between 0 and 1. The normalization keeps the features with
larger values from dominating the classifier training and hurt-
ing the accuracy. The D/W–training block finds the optimal
values of the tunable parameters of SVDE by performing grid
search along with 10-fold cross validation and selecting those
values for the tunable parameters that give highest accuracy.
Finally, the D/W–training block trains an SVDE classifier and
saves this classification model in a database for use at runtime.
If there are Ddoors and Wwindows, the D/W–training
block generates D+Wclassification modes. To enroll any
new door/window, all WiHS has to do is acquire its training
samples, train a classifier, and store it in the database. This
enables incremental deployment of WiHS instead of having to
enroll all doors/windows before WiHS becomes usable.
D. Runtime Identification of Door/Window
When the segmentation block reports that a door/window
event has happened, the D/W–Identification block takes the
denoised-stream of the event, generates a spectrogram, multi-
plies its y-axis with the duration of the event, and calculates
the mean and median of each frequency. To evaluate the de-
tected movement against the classification model of any given
door/window, the D/W–identification block selects the mean
and median magnitude values of the same frequencies that the
D/W–training block selected for the given door/window, and
obtains a feature vector of the same length that D/W–training
block used in generating the classification model of the given
door/window. Next, the D/W–identification block scales the
values in this feature vector using the same scaling factors that
the D/W–training block used during training, and evaluates
this normalized vector against the classification model of
the given door/window to calculate a likelihood value. It
calculates a likelihood value from all D+Wclassification
models, and declares the movement to be originating from that
door/window whose model returned the highest likelihood.
IV. OPE N/ CL OS E IDE NT IFI CATI ON
1) Intuition: Referring back to Fig. 2, we observe that the
variations introduced by door open in the denoised-stream are
opposite to those by door close, which is intuitive as door
open is the opposite movement of door close. Based on this,
we chose to employ the simple, yet very effective, approach
of dynamic time warping (DTW) to determine whether the
movement was an open event or a close event. DTW calculates
a score between any given pair of time-series: the more similar
the time-series, the smaller the DTW score.

2) Runtime Identification of Open/Close: As soon as the
D/W–identification block identifies the door/window that
moved, the O/C–identification block takes the denoised-stream
for that event and compares it with all training samples of the
opening and closing of the identified door/window and calcu-
lates a DTW score from each training sample. If the average
DTW score is smaller with the opening training samples, the
O/C–identification block declares the event to be door open,
otherwise door close. Note that the door/window does not have
to be completely opened or closed. The O/C–identification
block is able to distinguish between open and close movements
even when the door is partially moved because in case of
opening event, the DTW score with the training samples of the
opening of the identified door/window is still smaller than that
with the training samples of the closing of that door/window.
The only effect that the partial door/window movement has
is that the magnitude of the difference between the average
DTW score with opening training samples and the average
DTW score with closing samples decreases, which impacts
the final decision only minimally.
V. PERFORMANCE EVAL UATION
We implemented WiHS using commodity devices that in-
cluded TL-WR7500 WiFi access point (AP) and Lenovo X200
laptops, each with an Intel 5300 WiFi NIC and 3 omnidirec-
tional antennas. We used the tool presented in [17] to acquire
CSI measurements at a sampling rate of 300 samples/sec.
A. Test Houses
Fig. 12 shows the floor plans of our three test houses along
with the locations where we placed the WiFi AP (TX) and
the receiver laptops (RX). In each house we used one AP,
but a different number of receivers due to different sizes of
the houses. House (H) # 1 is a 1-bedroom apartment with 12
inch thick exterior walls made of brick and mortar and 6 inch
thick interior walls made of sheet rock. The doors are made
of 1.75 inch thick wood. The widths of doors D1 and D2 and
sliding windows W1 and W2 are 36, 28, 32, and 32 inches,
respectively. H # 2 is a 2-bedroom apartment with similar types
of walls as H # 1. D1, D2, and D3 are each 28 inch wide and
1.75 inch thick made of wood, and W1 to W4 are all 32 inch
wide. H # 3 is a 4-bedroom condo with 16 inch thick wooden
exterior walls and 6 inch thick interior sheet rock walls. D1 to
D6 are all 1.375 inch thick wooden with widths between 28
and 32 inches. W1 through W4 are all 48 inch wide. While
collecting data, in addition to our own AP, H # 1, 2, and 3
received signals from 10, 6, and 12 other APs, respectively,
of neighboring homes.
B. Overall Accuracy
1) Identification of the Door/Window that Moved: To
study the overall accuracy of WiHS in identifying which
door/window moved, we collected 60 open and 60 close events
for each door and window in each house in 4 sessions (15 open
and 15 close events per session). For each door, we asked our
volunteers to open and close the door from outside to imitate
an intrusion. Most volunteers took 2 to 4 seconds to open or
Bar
Kichen
RX1
RX2
TX
W2 W3
W4
D3
D4
D5
D2
W1
Bar
RX3
W2
Kichen
RX1
TX
RX3
RX2
W1
W3
W4
D1 D2
D3
W2
RX2
TX
RX1
W1
Closet
Closet
Closet
Closet
House # 1 House # 2 House # 3
D6
D1
D1
D2
Fig. 12. Floor plans of the three test houses
close a door by 90◦. We collected 480, 840, and 1200 samples
for H # 1, 2, and 3, respectively, over five weeks.
For this evaluation, we assigned the same label to the open
and close samples of any given door/window because we are
just identifying which door/window moved. For each house,
we took the 120 samples of each door/window and performed
10-fold cross validation. In each of the 10 classification
rounds, we used the method described in Secs. III-B and III-C
for training and Sec. III-D to evaluate the samples of test folds.
The black bars in Figs. 13(a) – 13(c) show the percentage of
120 samples of each door and window in the three houses that
WiHS identified correctly. We see from Fig. 13(a) that WiHS
identified all doors and windows with 100% accuracy in H #
1 due to the relatively smaller number of doors and windows.
The accuracies for H # 2 and 3 are slightly lower, but still over
95% across all doors and windows. Table I shows the confu-
sion matrix for house number 2. We have not included the
D1 D2 D3 W1 W2 W3 W4
D1 1 0 0 0 0 0 0
D2 0 .95 .04 0 .01 0 0
D3 0 .01 .99 0 0 0 0
W1 0 0 .01 .95 .01 .03 0
W2 0 0 0 0 .97 .03 0
W3 0 0 0 .03 .04 .93 0
W4 0 0 0 .01 0 .02 .97
TABLE I
CONFUSION MATRIX FOR HOUSE # 2
confusion matrix for
H # 1 as it is an
identity matrix and
for H # 3 due to the
similarity of obser-
vations as for H # 2.
We see in Tab. I that
WiHS made some mistakes for the doors and/or windows
that were very close to each other. This is intuitive because
different doors and windows affect CSI measurements dif-
ferently due to the differences in their positions. This slight
error does not have significant practical implications because
as long as the system points to a door/window that is close
to the actual door/window through which an intruder may be
breaking in, the home-occupants/law-enforcement still get the
correct information about where the intrusion started from.
2) Distinguishing between Open/Close Events: To study
this, we used the same data set as in Sec. V-B1 and performed
10 fold cross validation, this time for each window/door
separately. Figs. 13(a) to 13(c) plot the percentage of 60 open
and 60 close samples of each door and window identified
correctly in the three houses. We observe from these figures
that WiHS identified the open and close events of the majority
of doors/windows with at least 90% accuracy. The average
accuracy of WiHS in identifying the door open/close events
was 95.7%, 93.1%, and 91.9% for H # 1, 2, and 3, respectively.
3) Distinguishing Between Inside and Outside Movements
and Between Human and Door/Window Movements: To study
this, we collected CSI measurements for 60 instances of

D1 D2 W1 W2
80
85
90
95
100
Accuracy (%)
D/W Open Close
(a) House # 1
D1 D2 D3 W1 W2 W3 W4
80
85
90
95
100
Accuracy (%)
D/W Open Close
(b) House # 2
D1 D2 D3 D4 D5 D6 W1W2 W3W4
80
85
90
95
100
Accuracy (%)
D/W Open Close
(c) House # 3
Fig. 13. Percentage of samples of each door/window from which WiHS identified that door/window correctly and
from which WiHS correctly detected whether the door/window was opened or closed
0.05 0.1 0.15 0.2 0.25 0.3
20
40
60
80
100
Accuracy (%)
Non-walk Inside
Walk Outside
Walk Inside/D/W
Fig. 14. Segmentation accuracies of
WiHS with different values of α
walking outside H # 2, where in each instance a volunteer
walked outside along the path from W1 to D1 to W4 at a
distance of about 2 ft from the wall. We also collected CSI
measurements for 60 instances of walking inside H # 2, where
in each instance, the volunteer walked a random path either
between D1 and D2 or between W1 and W2. We further
collected CSI measurements for 30 open and 30 close events
of D1, 30 sitting down and 30 standing up events on the
sofa and the dining chair, 30 waving hand events, where the
volunteer randomly chose a different position in the house for
each waving hand event, and 30 eating/drinking events sitting
on the dining table chair.
To distinguish between the movement outside the house
and inside, as well as between the non-walking movements
and the walking/door/window movement inside the house, we
used the method in Sec. II-1. We used α= 0.25 to ignore
any movements that happen outside the house as well as any
non-walking movements that happen inside. We experimented
with various values of αranging from 0.05 to 0.30. To
distinguish between walking inside the house and door window
movements, we used the method described in Sec. II-2.
Fig. 14 plots the percentage of instances of walking outside
the house that WiHS ignored correctly using different values of
α. It also shows the percentage of non-walking instances inside
the house (i.e., sitting down, standing up, waving hand, and
eating/drinking) that WiHS ignored correctly. It further shows
the percentage calculated collectively over the 60 instances of
walking inside and the 30 open and the 30 close events of
D1 that WiHS detected correctly. We observe from this figure
that, WiHS is able to accurately ignore 100% of the walking
instances outside the house when α≥0.15 and 100% of the
non-walking instances inside the house when α≥0.25. We
also observe that WiHS is able to correctly detect the walking
and door/window movement 100% of the times at α= 0.05.
However, the accuracy decreases as αincreases. Therefore,
to keep a balance between correctly ignoring non-walking
movements and accurately detecting walking and door/window
movements, we have chosen α= 0.25 in WiHS, where WiHS
correctly ignores non-walking movements inside the house
with an accuracy of 100% and correctly detects the walking
10 15 20 25 30
Ratio
50
60
70
80
90
100
Accuracy (%)
Walk Inside Door
Fig. 15. Accuracy of distinguishing
between walk & door/window movt.
and door/window moveme-
nts with 96.7% accuracy.
Fig. 15 plots the percent-
age of walking instances in-
side the house that WiHS
correctly identified as walk-
ing and of 30 door open and
30 close instances that it correctly identified as door move-
ments for different values of R. We observe from this figure
that as Rincreases, the percentage of walking instances that
WiHS correctly recognizes increases. However, increasing R
too much increases incorrect identification of door open/close
instances. WiHS achieves highest accuracy when R= 20,
where it correctly identifies walking and door open/close
instances 95% and 96.7% of the times, respectively.
C. Impact of Real-World Characteristics
1) Number of WiFi Receivers: To evaluate the impact of the
number of WiFi receivers on the accuracy of WiHS, we used
the same data set that we used in Sec. V-B1. Just like in Sec.
V-B1, we measured the overall accuracy of WiHS in identify-
ing the door/window that moved, this time using CSI measure-
ments from 1 receiver, from 2 receiver, and from 3 receivers
separately. Fig. 16 plots the accuracy of WiHS using 1, 2, and
1 2 3
Number of receivers
80
85
90
95
100
Accuracy (%)
House # 1 House # 2 House # 3
Fig. 16. Accuracy vs. # of receivers
3 receivers in each house.
For H # 1, we do not show
accuracy with 3 receivers as
we only used 2 receivers
in it. We observe from this
figure that in each house, the
accuracy increases with the
increase in the number of receivers. We further observe that the
houses that are bigger in size, naturally, need more receivers
to achieve a target accuracy.
2) Speed of Opening/Closing the Door/Window: We col-
lected additional samples where we asked a volunteer to open
and close D1 and D2 in H # 2, where each door open and close
took approximately two seconds. We collected 15 samples of
door open and 15 of door close for each of the two doors. Next,
we collected same data for 4 second and then for 6 second
durations of door open and close. This way, we collected 180
new samples. To evaluate these samples, we used samples for
all 7 doors and windows that we collected in Sec. V-B1 and
trained the classifiers using the method described in Secs. III-B
and III-C. Next, we evaluated each of the 180 newly collected
samples using the trained classifiers and obtained a decision
for each sample. Our results showed that WiHS achieved
accuracies of 98.7%, 100%, and 95.1% in identifying the 60
samples of D1 and D2 with 2, 4, and 6 second opening/closing
times, respectively. This shows that WiHS is not significantly
impacted by the speed with which a door/window moves. The
high accuracy is the result of the approach described at the
end of Sec. III-B2 to normalize the speed of the moving
door/window before classification.

3) Status of the Other Doors and Windows in the House:
We collected 30 open and 30 close samples of D3 in H #
2. Before collecting each sample, we randomly changed the
state of the other doors and windows. We first evaluated the
accuracy in identifying that the moving door was D3 from each
of the 60 samples. For this, we used all samples for all 7 doors
and windows described in Sec. V-B1 and trained 7 classifiers.
When we were collecting samples for Sec. V-B1, we kept
all doors open and windows closed. Next, we evaluated the
60 newly collected samples against these 7 classifiers one
by one and obtained a decision for each sample as to which
class it belongs to. Second, we evaluated WiHS’s accuracy in
distinguishing between the open and close events of D3 from
each of the 30 open and close samples. For this, we used all
open and close samples for D3 in H # 2, described in Sec.
V-B1, and evaluated each of the 30 open and 30 close samples.
WiHS achieved 97.6% accuracy in identifying that D3 was
moving, which is similar to the 98.5% accuracy for D3 in Fig.
13(b). Our results further showed that WiHS achieved 91.5%
accuracy in correctly identifying whether D3 was opened or
closed, which is similar to the 92.2% accuracy in Fig. 13(b)
. These observations show that the open/close status of other
doors and windows does not have any noticeable impact on the
accuracy of WiHS. This is because WiHS relies on measuring
the changes in the denoised-streams caused by the movement
of a door/window. The multi-paths reflected from doors that
are already open or close are part of Hs(f)in Eq. (1) and
thus do not contribute any frequencies to the denoised-stream.
4) Partial Opening/Closing of Door/Window: To evaluate
how partial opening/closing of a door (i.e. by >45◦and <90◦)
or window (i.e. by greater than half way and less than all the
way) impact the accuracy, we asked a volunteer to partially
open and close D1 and W2 in H # 2. We collected 30 samples
each for door open, door close, window open, and close and
evaluated them against the models trained in Sec. V-B1. WiHS
achieved accuracies of 100% and 96.7% in identifying that the
moving entity was D1 and W2, respectively. This accuracy is
very similar to what we saw in Tab. I. This shows that WiHS
can accurately identify the moving door/window even when
the door/window is not fully opened/closed.
5) Presence and Movements of Other People: To study how
the presence and movements of occupants impact WiHS, we
collected samples from D1 in H # 2 in 5 scenarios (S1 to S5),
where in each scenario, we collected 30 open and 30 close
samples. In S1, an occupant was sleeping in room 1 while we
collected samples. In S2, an occupant used his phone sitting
on the bed in room 1 while we collected samples. In S3, an
occupant practiced yoga in room 2, in S4, an occupant walked
in the vicinity of D1 but outside the house, and in S5, an
occupant cooked in the kitchen while we collected samples.
We evaluated the 60 samples of D1 in each scenario in the
same way as we evaluated the 60 samples of D3 in Sec. V-C3.
Fig. 17 plots the accuracy of WiHS in identifying that D1 was
moving and if it was opened or closed. We see that for S1 to
S4, WiHS achieved the same 100% accuracy in identifying D1
as we saw in Sec. V-B1. However, in S5, the accuracy dropped
S1 S2 S3 S4 S5
40
60
80
100
Accuracy (%)
Door Movt. Open Close
Fig. 17. Impact of presence & move-
ment of people on WiHS’s accuracy
to 58% due to a lot of dy-
namic activities by the oc-
cupant such as opening and
closing refrigerator, cutting
vegetable, sitting and stand-
ing, etc. Similarly, in deter-
mining whether the door was
opened or closed, the accuracies dropped in the scenarios
where movements of the occupant were more prominent. We
emphasize that while the accuracy of WiHS dropped in S5, its
accuracy in detecting that a movement has occurred remained
unaffected. WiHS still raises an alarm; its just that in such a
scenario, it does not accurately determine which door moved.
Thus, WiHS is best useful when the occupants have all gone
outside, i.e., the away-mode, or when they are not performing
any major activities (such as going to bed at night), i.e., the
stay-mode, and need the home to be monitored for intrusions.
These two times, i.e., when occupants are not home or at night
time, are usually the two times where users of conventional
security systems arm them to perform security monitoring.
VI. RE LATE D WOR K
Wireless signal based. Several prior works use received signal
strength to determine the presence of humans in the given
environment (mostly at a room level) [18], [19]. Recently,
Wu et al. proposed DeMan that utilized temporal variations in
CSI measurements to detect a moving person and breathing-
induced periodic patterns in CSI measurements to detect a
stationary person [20]. We emphasize that this line of work
does not overlap with the movement segmentation component
of WiHS because the focus of prior work is only to determine
whether one or more humans are present in the coverage area
of a WiFi transceiver, whereas the focus of the movement
segmentation component of WiHS is not only to detect move-
ments in a given house but also to determine whether the
movement was inside the house or outside, or whether the
movement was from a human or doors/windows.
Another, relatively more recent, line of work uses CSI
measurements to recognize human gestures and activities [11],
[14], [15]. Wang et al. modeled the impact of human activities
on CSI measurements and used this model to develop a single
user activity recognition system [14]. Venkatnarayan et al.
extended the work in [14] and modeled the impact of activities
of multiple simultaneously moving humans on CSI measure-
ments and used this model to develop a multi-user activity
recognition system [11]. Xi et al. analyzed the relationship
between the CSI variation and the number of moving people
to count people in a crowd [21]. E-eyes used distributions of
the CSI measurements to recognize activities such as cooking,
bathing etc. [22]. SIED detected human intrusion using the
variance in CFR [23]. Li et al. proposed AR-Alarm that
distinguishes human movements from movements of everyday
objects, such as curtains [24]. While SIED and AR-Alarm
detect human motion inside the house, they can not distinguish
between movements inside and outside the house and further
do not detect whether a door/window was opened/closed.

Sensor based. Many other sensors such as accelerometers,
passive infrared (PIR) motion sensors, vibration sensors,
barometers, and even cameras have been used to detect
door/window/human movements. Commercial products such
as [25], [26] can detect an intruder by sensing vibration when
a door or window is opening. In [27], an energy-harvesting
vibration sensor was proposed that detects door open events
when attached to the door. Patel et al. utilized pressure
sensors installed in the air conditioning ducts to detect pressure
variation caused by door open/close events [28]. Wu et al.
[29] used barometer in a smartphone to detect door open/close
events anywhere inside a building. These pressure sensor and
barometer based approaches usually do not distinguish across
different doors, especially if the doors are of the same size.
Today’s commercial home security systems use magnetic
contact sensors to detect door/window open/close events and
PIR sensors to detect human motion. As me mentioned in the
introduction, while these provide deterministic intrusion de-
tection, WiHS is not competing with them, rather is providing
a solution for homes that otherwise do not have a security
system at all. In many ways, WiHS is actually complimentary
to these commercial systems. For example, PIR sensors detect
human motion only in the line of sight, while WiHS can detect
any motion in non line of sight as well. Similarly, due to
their cost, conventional systems deploy contact sensors only
at the main entrance. While they monitor the main entrance
of a house, WiHS can monitor the door movements inside the
house, which is important for post-incident analysis (e.g., to
determine what locations inside the house an intruder visited).
VII. CONCLUSION AND FUTURE WOR K
In this paper, we proposed WiHS, a WiFi based home
security system that accurately performs the three primary
monitoring tasks of the typical home security systems. The
key technical novelty of WiHS lies in developing the theoret-
ical understanding of the impact of the movements of doors
and windows on CSI measurements. The key technical depth
of WiHS lies in the techniques that it uses to 1) distinguish
between the movements inside and outside the house, 2)
distinguish between walking, non-walking, and door/window
movements, 3) identify the door/window that has moved, and
4) determine whether the moving door/window opened or
closed. The results from an extensive evaluation of WiHS on
commodity WiFI devices in three different houses show that
WiHS detects intrusions with over 95% accuracy.
WiHS also has its share of limitations. For example, WiHS
works best in a single story house. For a multi-story house,
each floor will need its own instance of WiHS deployed
using the hardware on that floor. WiHS also requires that
the locations of transceivers stay fixed. This means that to
deploy it on existing equipment, the house will need WiFi
receivers that do not change their position, such as a desktop
computer. One could also achieve this by, for example, using
a few raspberry Pis and attaching them to various wall-sockets
in the house. This, however, entails some monetary cost. We
plan to address these aspects in our future work.
REFERENCES
[1] S. Lee, “The impact of home burglar alarm systems on residential
burglaries.”
[2] A. Tseloni, R. Thompson, and N. Tilley, Reducing Burglary. 2018.
[3] www.samsung.com/us/smart-home/smartthings/sensors
[4] www.adt.com/door-window-alarms
[5] J. J. Roth, “The role of perceived effectiveness in home security choices,”
Security Journal, vol. 31, no. 3, pp. 708–725, 2018.
[6] Burglary crime statistics and facts. www.alarms.org/burglary-statistics/
[7] www.chinairr.org/report/R03/R0304/201808/27-271833.html
[8] E. Soltanaghaei, A. Kalyanaraman, and K. Whitehouse, “Multipath
triangulation: Decimeter-level wifilocalization and orientation with a
single unaided receiver,” in MobiSys, 2018, pp. 376–388.
[9] K. Qian, C. Wu, Y. Zhang, G. Zhang, Z. Yang, and Y. Liu, “Widar2. 0:
Passive human tracking with a single wi-filink,” MobiSys, 2018.
[10] A. Virmani and M. Shahzad, “Position and orientation agnostic gesture
recognition using wifi,” in MobiSys 2017.
[11] R. H. Venkatnarayan, G. Page, and M. Shahzad, “Multi-user gesture
recognition using wifi,” in MobiSys 2018.
[12] M. Shahzad and S. Zhang, “Augmenting user identification with wifi
based gesture recognition,” IMWUT, vol. 2, no. 3, 2018.
[13] K. Ali, A. X. Liu, W. Wang, and M. Shahzad, “Keystroke recognition
using wifisignals,” in MobiCom, 2015.
[14] W. Wang, A. X. Liu, M. Shahzad, K. Ling, and S. Lu, “Understanding
and modeling of wifisignal based human activity recognition,” in
MobiCom, 2015.
[15] W. Wang, A. X. Liu, and M. Shahzad, “Gait recognition using wifi
signals,” in IMWUT, 2016.
[16] Y. Xu, W. Yang, J. Wang, X. Zhou, H. Li, and L. Huang, “Wistep:
Device-free step counting with wifisignals,” Proc. ACM Interact. Mob.
Wearable Ubiquitous Technol., vol. 1, no. 4, pp. 172:1–172:23, Jan.
2018.
[17] D. Halperin, W. Hu, A. Sheth, and D. Wetherall, “Tool release: Gathering
802.11 n traces with channel state information,” ACM SIGCOMM
Computer Communication Review, vol. 41, no. 1, pp. 53–53, 2011.
[18] M. Moussa and M. Youssef, “Device-free passive detection in real
environments,” in Pervasive Computing and Communications, 2009.
PerCom 2009. IEEE International Conference on. IEEE, 2009, pp.
1–6.
[19] J. Yang, Y. Ge, H. Xiong, Y. Chen, and H. Liu, “Performing joint learn-
ing for passive intrusion detection in pervasive wireless environments,”
in INFOCOM, 2010 Proceedings IEEE. IEEE, 2010, pp. 1–9.
[20] C. Wu, Z. Yang, Z. Zhou, X. Liu, Y. Liu, and J. Cao, “Non-invasive
detection of moving and stationary human with wifi,” IEEE Journal
on Selected Areas in Communications, vol. 33, no. 11, pp. 2329–2342,
2015.
[21] W. Xi, J. Zhao, X.-Y. Li, K. Zhao, S. Tang, X. Liu, and Z. Jiang,
“Electronic frog eye: Counting crowd using wifi,” in Infocom, 2014
proceedings IEEE. IEEE, 2014, pp. 361–369.
[22] Y. Wang, J. Liu, Y. Chen, M. Gruteser, J. Yang, and H. Liu, “E-
eyes: device-free location-oriented activity identification using fine-
grained wifisignatures,” in Proceedings of the 20th annual international
conference on Mobile computing and networking. ACM, 2014, pp.
617–628.
[23] J. Lv, W. Yang, L. Gong, D. Man, and X. Du, “Robust wlan-based indoor
fine-grained intrusion detection,” in 2016 IEEE Global Communications
Conference (GLOBECOM), Dec 2016, pp. 1–6.
[24] S. Li, X. Li, K. Niu, H. Wang, Y. Zhang, and D. Zhang, “Ar-alarm:
An adaptive and robust intrusion detection system leveraging csi from
commodity wi-fi,” in Enhanced Quality of Life and Smart Living, 2017.
[25] www.zwaveproducts.com/shop/z-wave-security/
z-wave-door- and-window-sensors/door-window-sensor-gen-5
[26] www.vigilantpps.com/products/mace-slimline-wireless- door-window-sensor
[27] B. Campbell and P. Dutta, “An energy-harvesting sensor architecture
and toolkit for building monitoring and event detection,” in Proceedings
of the 1st ACM Conference on Embedded Systems for Energy-Efficient
Buildings. 2014.
[28] S. N. Patel, M. S. Reynolds, and G. D. Abowd, “Detecting human
movement by differential air pressure sensing in hvac system ductwork:
An exploration in infrastructure mediated sensing,” in International
Conference on Pervasive Computing. 2008.
[29] M. Wu, P. H. Pathak, and P. Mohapatra, “Monitoring building door
events using barometer sensor in smartphones,” in UbiComp, 2015.