Translation Resilient Opportunistic WiFi Sensing
Mohammud J. Bocus
1
, Wenda Li
2
, Jonas Paulavicius
1
, Ryan McConville
1
, Raul Santos-Rodriguez
1
,
Kevin Chetty
2
and Robert Piechocki
1
1
Department of Electrical and Electronic Engineering, University of Bristol, BS8 1UB, UK.
2
Department of Electronic and Electrical Engineering, University College London, UK
Abstract—Passive wireless sensing using WiFi signals has be-
come a very active area of research over the past few years. Such
techniques provide a cost-effective and non-intrusive solution
for human activity sensing especially in healthcare applications.
One of the main approaches used in wireless sensing is based
on fine-grained WiFi Channel State Information (CSI) which
can be extracted from commercial Network Interface Cards
(NICs). In this paper, we present a new signal processing
pipelines required for effective wireless sensing. An experiment
involving five participants performing six different activities was
carried out in an office space to evaluate the performance of
activity recognition using WiFi CSI in different physical layouts.
Experimental results show that the CSI system has the best
detection performance when activities are performed half-way
in between the transmitter and receiver in a line-of-sight (LoS)
setting. In this case, an accuracy as high as 91% is achieved while
the accuracy for the case where the transmitter and receiver are
co-located is around 62%. As for the case when data from all
layouts is combined, which better reflects the real-world scenario,
the accuracy is around 67%. The results showed that the activity
detection performance is dependent not only on the locations of
the transmitter and receiver but also on the positioning of the
person performing the activity.
I. INTRODUCTION
WiFi has experienced rapid growth due to an increasing
number of wireless devices. Its success is also due to the
multiple-input multiple-output (MIMO) technology which has
boosted the throughput to meet the growing demand of wire-
less data traffic. Using the Orthogonal Frequency Division
Multiplexing (OFDM) physical layer waveform, WiFi stan-
dards such as IEEE 802.11n provide channel state information
(CSI) for each received packet. The WiFi signal emitted from
the commodity routers interacts with surrounding objects and
humans which are present in the propagation environment.
Furthermore, CSI captures the changes in wireless character-
istics of the environment. Although CSI is included in WiFi
since IEEE 802.11n, it is typically not reported by off-the-shelf
WiFi Network Interface Cards (NICs). Therefore, dedicated
tools need to be used to extract the CSI from the NICs. For
example, the Atheros [1] and Linux 802.11n [2] CSI tools
and custom modified firmware and open source Linux wireless
drivers at the end-user equipment, CSI has been successfully
extracted from specific NICs such as Intel 5300 (IWL5300)
and some Atheros chipsets.
Using WiFi signal for human sensing has gained much
attention over the past decade. Compared to other technologies
such as cameras [3], [4] and wearable sensors [5], [6], WiFi
based sensing technology is considered as an ideal solution
for future smart homes and healthcare residences since it is
not intrusive, provides a wider coverage, is not sensitive to
lighting conditions and the users do not need to wear any
uncomfortable devices, making this technology completely
passive.
The CSI obtained from commodity WiFi devices can be
used for multiple indoor monitoring purposes like activity
[7]–[9] and gait recognition [10], fall detection [11], [12],
gesture [13], [14] and sign recognition [15], human/intrusion
detection [16] , crowd counting [17], among other applications.
Note that there are schemes that use specialised hardware
such as Universal Software Radio Peripheral (USRP) [14]
and directional antennas [8] to obtain fine-grained signal
measurements from WiFi signals. For instance, WiHear [8]
makes use of directional antennas to obtain CSI variations
caused by the movement of the lips for recognising spoken
words. WiHear uses directional antennas to reduce the noise
in the CSI data to achieve acceptable accuracy. WiSee [14]
uses USRP to capture the OFDM signals and measures the
Doppler shift in the signals caused by the reflection from
human bodies to recognise nine gestures with an accuracy
of 95%. E-eyes [9] uses CSI histograms as fingerprints for
recognising nine daily human activities such as brushing teeth,
taking shower, washing dishes, walking, etc., with an accuracy
of 92%. It detects the activities of a single person using
location-oriented time domain features. CARM [7], on the
other hand, requires no specialised hardware to achieve high
activity recognition accuracy (≈96%). It classifies activities
performed by a single person such as falling down, walking
and sitting down using time-frequency features such as torso
and limb velocities by characterising the changing speed of
the reflected path length on subcarrier amplitudes. WiFiU [10]
extends CARM to recognise human gait from limb and torso
velocities. WiFiU can detect a walking human subject at a
range as long as 14 meters with an accuracy of 92%.
Compared to previous studies such as [18], [19], this work
makes the following contributions:
• We leverage the use of commodity WiFi devices with no
specialised hardware for activity recognition in different
physical layouts, covering both Line-of-Sight (LoS) and
non LoS scenarios.
• The participants performed the activities in a random
fashion or different orientations with respect to the trans-
mitter/receiver (not controlled experiment). This is more
representative of the real-world scenario.
CONFIDENTIAL. Limited circulation. For review only.
Manuscript 1534 submitted to 2020 25th International Conference
on Pattern Recognition. Received April 15, 2020.
Fig. 1. Block diagram of signal processing techniques applied to CSI data.
• The experimental results evaluate the performance of the
WiFi CSI system and identify the layout and coverage
sensitivities. The results provide a benchmark for the ex-
pected accuracy in different physical transmitter-receiver
geometry at different positions.
In this paper, we focus on activity recognition using WiFi
CSI. The CSI data was collected alongside video recording
during the whole experiment to provide ground truth labels.
Both the CSI data and video were timestamped using the
Network Time Protocol (NTP) Servers for synchronisation
purposes. We demonstrate the detection result of the system for
six different activities performed by five participants namely,
sitting down, standing up, walking, laying down, standing from
the floor and picking up an object from the ground.
The rest of the paper is organised as follows. An overview
of the WiFi CSI system is given in Section II. The signal
processing for the CSI system is described in Section III.
The system implementation and experiment description are
provided in Section IV. The activity recognition performance
is evaluated in Section V. Finally, conclusions are drawn at
the end of this paper.
II. OVERVIEW OF WIFI CSI
OFDM is a widely used technology in many WiFi stan-
dards including IEEE 802.11 a/g/n/ac. In an OFDM system,
the bandwidth is shared among multiple overlapping but
orthogonal subcarriers (subchannels) and therefore they do not
interfere with each other. Since the subcarriers have a small
bandwidth, they experience only flat fading. Thus, OFDM
provides robustness in a frequency-selective fading wireless
channel. For a WiFi system with MIMO-OFDM capability, in
each packet the CSI is obtained as a 3-dimensional (3D) matrix
with n
t
× n
r
× N complex values, where n
t
is the number of
transmit antennas, n
r
is the number of receive antennas and
N is the number of subcarriers. Assuming a narrowband flat-
fading channel, the frequency domain input-output relation in
a MIMO-OFDM system with N subcarriers can be represented
as
y
k
= H
k
x
k
+ z
k
, k = 1, 2, · · · , N (1)
where
x
k
= [x
1
, x
2
, · · · , x
n
t
]
T
, (2)
y
k
= [y
1
, y
2
, · · · , y
n
r
]
T
, (3)
z
k
= [z
1
, z
2
, · · · , z
n
r
]
T
, (4)
are the transmitted signal, received signal and noise vectors,
respectively. For a MIMO system, H
k
has a dimension of
N
r
× N
t
and is represented as
H
k
=
h
1,1
h
1,2
· · · h
1,n
t
h
2,1
h
2,2
· · · h
2,n
t
.
.
.
.
.
.
.
.
.
.
.
.
h
n
r
,1
h
n
r
,2
· · · h
n
r
,n
t
, (5)
where h
i,j
is the complex-valued channel coefficient between
the jth transmit antenna and ith receive antenna. To recover
the transmitted signal in a wireless medium, the receiver needs
to estimate the channel. In this regard, a training sequence
which is known by both the transmitter and receiver is sent
in each packet to obtain the channel estimate. This process is
often referred to as channel sounding. For a MIMO-OFDM
system, the channel estimate (i.e., CSI) is a matrix consisting
of complex values for each subcarrier, as in Equation (5). The
equaliser uses the CSI to reverse the effects of the channel such
as multipath propagation, attenuation, phase shift, scattering,
etc., to recover the transmitted signal. In order to have an
insight on the equalisation process, consider a simple single-
input single-output (SISO) OFDM system, i.e., n
t
= n
r
= 1.
Let the received frequency domain signal, following the Fast
Fourier transform (FFT) operation at the OFDM receiver, be
denoted as
Y = HX + Z. (6)
In 802.11n, out of the 64 subcarriers that occupy the 20
MHz bandwidth, 4 subcarriers are used as pilots and 52
as data-carrying subcarriers. The remaining subcarriers are
unused (they serve as DC/null subcarriers). As for the 40
MHz bandwidth, out of a total of 128 subcarriers, only 114
subcarriers are used, out of which 6 serve as pilots. The
channel estimate at each pilot position can be obtained as
ˆ
H
p
(k) =
Y
p
(k)
X
p
(k)
, p = 1, 2, · · · , N
p
(7)
where p is the pilot index, N
p
is the total number of pilot
subcarriers,
ˆ
H
p
(k) is estimated the channel frequency response
(CFR) at the pilot subcarriers, X
p
and Y
p
are the transmitted
and received pilot signals, respectively. In order to estimate the
channel at the data subcarriers, the receiver performs channel
interpolation.
CONFIDENTIAL. Limited circulation. For review only.
Manuscript 1534 submitted to 2020 25th International Conference
on Pattern Recognition. Received April 15, 2020.
III. SIGNAL PROCESSING FOR ACTIVITY SENSING
CSI measurements in the time domain capture the changes
in the wireless signal due to the latter’s interaction with sur-
rounding objects or human activities and the observed patterns
can be used for various purposes. Different WiFi sensing
applications have specific requirements in terms of their signal
processing techniques and classification/estimation algorithms.
This section presents the signal processing techniques applied
to the raw CSI data, as illustrated in Fig. 1.
A. Phase Calibration
In WiFi systems, the true phase information in the raw
CSI measurements is corrupted due to hardware errors. These
include Sampling Frequency Offset (SFO), Packet Detection
Delay (PDD) and Carrier Frequency Offset (CFO) [1]. In
this work, the phase is calibrated as in [20] where a linear
transformation is applied to the raw phase data to eliminate
the phase offset. The measured phase
ˆ
φ
i
of the ith subcarrier
be expressed as:
ˆ
φ
i
= φ
i
− 2π
k
i
N
δt + β + Z, (8)
where φ is the true phase, β is the phase offset due to CFO,
δt is the timing offset between the transmitter and receiver,
k
i
is the index of the ith subcarrier and Z is noise. In the
Intel 5300 CSI tool, i ∈ {1, 30} and N is the FFT size. For
example, N = 64 for a 20 MHz WiFi channel in IEEE 802.11
a/g/n. The terms δt, β and Z make it difficult to obtain the
true phase from WiFi NICs. The phase obtained from the raw
CSI measurements is corrected by first unwinding it and then
applying a linear transformation. The main idea is to remove
the terms δt and β by considering the phase across the whole
frequency band [20].
B. Noise Reduction
Since the raw CSI data is too noisy to be directly used in
machine learning algorithms for activity sensing, we use the
Discrete Wavelet Transform (DWT) technique to filter out in-
band noise and preserve the high frequency components, thus
avoiding significant distortion to the signal. DWT-based noise
filtering consists of transforming the signal into the wavelet
domain whereby the signal is divided into several frequency
levels called wavelets. More specifically, the signal is passed
through a set of high pass and low pass filters at each level.
The output from the high pass and low pass filters provides
the detailed and approximation coefficients, respectively [11].
The detailed coefficients in the first level contains information
about the noise and the abrupt changes caused by human
activity. Therefore, the detailed coefficients in the first level
are used to compute a threshold. The latter is adapted for
lower wavelets and the noise is removed in all levels without
introducing significant distortion to the signal. In addition to
DWT denoising, 1-D median filtering is also applied to the
signal to remove any undesired transients in the signal which
are not caused by human motion.
C. Data Compression
The raw CSI measurements were collected on a device
equipped with the Intel 5300 NIC. Considering one transmit
and three receive antennas, we obtain 1 × 3 × 30 = 90
complex CSI values in each packet. The packet rate was set
at 1 kHz. This results in a significant amount of data that
needs to be processed and which will serve as input to a
learning algorithm for classification. Therefore, dimensionality
reduction is required to reduce the computational complexity.
In this work, the PCA dimensionality reduction and denoising
technique has been used. PCA is used to identify the time-
varying correlations between the CSI streams which are then
optimally combined to extract components that represent the
variation caused by human activities. The number of principal
components (PCs) is empirically selected to achieve a good
trade-off between classification performance and computa-
tional complexity [10]. Following DWT denoising, the first
two or three PCs are, on average, sufficient to capture most
of the variance (≈70-80%) in the original 30 subcarriers [11].
Similar to [10], in the CSI system we extract the first six PCs.
However, the first one is safely discarded since it contains
noise due to reflection from stationary objects like furniture,
walls etc., and therefore discarding it will not result in any
loss of information [7], [10], [17]. Therefore, only the next
five PCs are retained for further processing.
D. Moving Variance Segmentation
In this step, the meaningful CSI variation caused by a human
activity is segmented using the Moving Variance Segmentation
(MVS) [21] method. The main idea lies in computing the mov-
ing variance along the entire CSI stream in a stepwise manner.
More specifically, in each step, a moving variance is calculated
over a sliding window of length L across neighbouring CSI
samples (packets), and the window is centred about the CSI
sample in the current position. A high moving variance value
represents significant variation in the CSI stream due to human
motion, whereas a low value represents only slight fluctuations
as in a stable environment. For a CSI stream consisting of M
packets, the moving variance is defined as
CSI
mv
=
M
X
m=1
"
1
L − 1
L
X
l=1
|CSI
l∈L
− µ
2
|
#
, (9)
µ =
1
L
L
X
l=1
CSI
l
,
where µ and l denote the mean and packet number in the
sliding window of length L, respectively, and m is the current
sample position in the CSI stream. An example of the CSI
stream for the standing activity and its corresponding moving
variance stream is shown in Fig. 2, where a sliding window of
length L = 100 is considered. The sliding window length L
needs to be chosen empirically to obtain the best results [21].
As can be observed in Fig. 2(b), the meaningful variations
due to human motion become greater in magnitude while
the insignificant fluctuations (not caused by human activity)
CONFIDENTIAL. Limited circulation. For review only.
Manuscript 1534 submitted to 2020 25th International Conference
on Pattern Recognition. Received April 15, 2020.
0 1000 2000 3000
20
25
30
35
40
(a)
0 1000 2000 3000
0
5
10
15
20
(b)
Fig. 2. MVS for standing activity (a) CSI stream (b) Corresponding moving
variance sequence.
become smaller. Therefore, the start and end points of an
activity can be easily identified and thus segmentation can be
performed to improve the system’s performance.
E. Spectrogram Generation
The CSI is highly sensitive to the surrounding environment
and signal reflections from the human body result in different
frequencies when performing different activities. These fre-
quencies can be distinguished in the time-frequency domain
(spectrogram) by applying the Short-time Fourier Transform
(STFT) to the segmented PCA-denoised signal. Basically, the
STFT applies a sliding window to obtain equally-sized seg-
ments of the signal and then performs FFT on the samples in
each segment. The spectrogram has three dimensions, namely,
time, frequency, and FFT amplitude. The Doppler spectrogram
from STFT identifies the change of frequencies over time.
The window size for FFT determines the trade-off between
frequency and time resolutions. For instance, a larger window
size results in a higher frequency resolution but lower time
resolution. The spectrograms are generated from the five PCs
which are then averaged to obtain the final spectrogram.
Fig. 3. Experiment layouts.
IV. SYSTEM IMPLEMENTATION & EXPERIMENT
DESCRIPTION
A. System Implementation
The Intel 5300 [2] NIC has been used in our CSI system.
Recall that the Intel 5300 Linux CSI tool extracts the CSI
measurements from 30 out of 56 subcarriers for each transmit-
receive antenna pair. The CSI data and video ground truth
were synchronised to an external NTP server to provide
timestamped measurements. The transmitter (TX) was a TP-
Link access point (AP) and the receiver (RX) was an Intel
Next Unit of Computing (NUC) device equipped with the Intel
5300 NIC, from which CSI was extracted and stored for off-
line processing. The CSI was collected over 3 antennas (30
subcarriers each) on the receiver side (NUC) in the 2.4 GHz
band (20 MHz bandwidth) by pinging the AP at a rate of 1
kHz. This rate was selected to capture noticeable changes or
patterns in the time domain signal which are caused by human
motion.
B. Experiment Layouts
The CSI data collection was carried out in an office space.
The experiment layouts are shown in Fig 3. The monitoring
area was approximately 8 m × 6 m with computers, chairs,
cabinets and desks in the surroundings. The receiver (NUC)
location was kept unchanged while the WiFi AP (TX) was
moved in each layout as per Fig. 3, where 3 different layouts
were considered. In Layout 1, the TX and RX were facing
each other (LoS) while in Layout 2, the TX and RX were
at 90
0
to each other. Finally, in Layout 3 the TX and RX
were co-located (placed next to each other). 9 testing positions
were used during the experiments and they were separated by
1.5 m from each other. These points are used to evaluate the
effect of the system geometry on the activity classification
accuracy. In this study, we conducted 6 activities, namely,
walking, standing from a chair, sitting on a chair, laying down
on the floor, standing from the floor and picking a small object
from the floor. The activity descriptions are given in Table I. 5
participants of different age groups performed the 6 activities,
CONFIDENTIAL. Limited circulation. For review only.
Manuscript 1534 submitted to 2020 25th International Conference
on Pattern Recognition. Received April 15, 2020.
(a) (b) (c)
(d) (e) (f)
Fig. 4. Layout 2 spectrograms: (a) walking, (b) sitting, (c) standing, (d) laying down, (e) standing from floor, (f) picking up.
TABLE I
EXPERIMENT DESCRIPTION.
Activity Description
1 walking walking along positions 1-2-3, 4-5-6, 7-8-9, 1-4-7,
2-5-8 and 3-6-9
2 sitting sitting on a chair at positions 2,4,5,6,8
3 standing standing from a chair at positions 2,4,5,6,8
4 laying
down
laying down on the floor at positions 2,4,5,6,8
5 standing
from
floor
standing from the floor at positions 2,4,5,6,8
6 picking
up
picking up objects from the floor at positions
2,4,5,6,8
one at a time, at the various positions for each layout. Note that
the activities were performed in a random fashion or different
orientations in a natural way, as would be the case in the real-
world scenario.
V. EXPERIMENTAL RESULTS
In this section, the performance of activity recognition in
different layouts is presented. Fig. 4 presents the spectrograms
for the 6 activities in the Layout 2 configuration. As observed
in Fig. 4, the activities that involve rapid body motion such
as walking have high energy in the higher frequencies in the
spectrogram. It is to be noted that the spectrograms contain
information about instantaneous frequency (positive) changes
but not motion direction, unlike radar systems. However, the
frequency-domain features obtained from a CSI spectrogram
are usually enough to identify a given activity. In this work,
instead of manually identifying a set of features, we use a
2D Convolutional Neural Network (CNN) for classification
purposes. The architecture of the 2D CNN is shown in Fig.
TABLE II
CLASSIFICATION PERFORMANCE FOR EACH LAYOUT.
Layout Precision Recall F1-score Accuracy
1 90.0% 89.5% 89.1% 90.8%
2 73.9% 74.4% 73.8% 75.7%
3 62.6% 61.9% 61.0% 61.5%
1, 2, 3 67.5% 66.3% 66.0% 67.3%
5. The network consists of a convolutional layer with 64
filters and 2×2 kernel size and the rectifier as the activation
function. Th next layer is the max-pooling layer with a stride
of 2. The output from the max-pooling layer is flattened to
create a single long feature vector (1D). The latter is fed
to the first fully-connected layer consisting of 64 filters and
using rectifier as the activation function. The second fully-
connected layer consists of 32 filters with the same activation
function as the previous layer. Finally, a softmax layer is used
for classification of the 6 classes of activities. 80% of the
dataset was randomly chosen and used for training while the
remaining 20% was used for testing. As shown in Fig. 6(a), the
overall accuracy for the case when data from all layouts are
combined is 67%. The reason for the low accuracy is because
the LoS and non-LoS data from the different layouts are mixed
together and these have different Doppler signatures. The best
classification result is obtained for activity 1 (walking) which
is more than 90%. This is because the walking activity has
higher Doppler shifts than other activities irrespective of the
direction or layout. The second best result is observed for
activity 6 (picking up) which is more than 70%. The incorrect
predictions most happen between the pair of activities such as
sitting and standing from a chair, laying down and standing
from the floor. The confusion matrix for Layout 1 (LoS) is
CONFIDENTIAL. Limited circulation. For review only.
Manuscript 1534 submitted to 2020 25th International Conference
on Pattern Recognition. Received April 15, 2020.
