5072 IEEE SENSORS JOURNAL, VOL. 20, NO.

9, MAY 1, 2020

Deep Learning Radar Design for

Breathing and Fall Detection
Abhijit Bhattacharya , Student Member, IEEE, and Rodney Vaughan, Life Fellow, IEEE

Abstract —The automated detection of people having a

fall is particularly important for the elderly in indoor care
situations. Privacy concerns, and regulations that prohibit
cameras in indoor environments, mean that optical sensing
must give way, at least in some situations, to less explicit
sensing such as radar. Currently, fall detection using radar
Doppler signatures has limitations. We demonstrate a radar-
based technique that detects breathing and other movements
seamlessly, and can detect a fall after it has happened,
i.e., even when the person is static. Using a low-cost, ceiling-
mounted radar at low microwave frequencies (sub-6 GHz), our
results show it is possible to remotely localize a person within
a few cm. The sensor system includes a small neural network
model that can distinguish a person from other moving objects in an indoor environment. This can reduce false alarms of
the fall detection and hence improve system reliability in real-world deployment. In our experiments, the neural network
differentiates a person from a pet (an example of a complex moving entity) with an accuracy exceeding 95%. This sensor
system demonstration paves a way forward for general indoor fall detection, extending to the well-being of our elderly
through real-time, ongoing monitoring of their breathing and other activities of daily living.
Index Terms — Radar sensor, breathing, fall detection, patient monitoring, deep convolutional neural network, IoT.

I. I NTRODUCTION of remote monitoring of vital-signs (breathing and heartbeat)

using radars [12], [13], and studied elderly fall detection from
R ADAR sensors are appearing in everyday life alongside
conventional commodity sensors such as cameras, and
infrared and ultrasonic technology. Radar technology is envi-
Doppler signatures to distinguish fall from daily activities
[14], [15]. Fall detection from Doppler signatures has seen
ronmentally robust, for example working in unilluminated limited success for two main reasons. Firstly, due to the
settings, and integrated circuit advances have enabled small, complex and varied nature of falls, it has proved difficult to
affordable short-range radars. They operating at frequencies differentiate from some other normal activities when observing
from few GHz to hundreds of GHz (mmwaves) for very short- from different angles [16]. Secondly, it is hard to collect real-
range sensing. This has made radar a choice for many new world fall data, which forces researchers to use data gathered
applications such as: collision avoidance in cars; adaptive con- from young volunteers, and this limits the system performance
trol of autonomous vehicles and drones ( [1], [2]); structural in real-world implementations.
monitoring; airport security ( [3], [4]); gesture recognition This work contributes three new aspects to the study of
[5], [6]; and so on. Patient monitoring is another area of radar-based fall detection. (i) Fall detection and vital-sign
interest in recent years [7]–[9]. For such human monitoring, monitoring have largely been treated in the literature as
conventional wearable sensors [10] are often cumbersome, and two separate problems. Here we locate the person through
people prone to forget to wear them. Cameras [11] are seen as movement as well as breathing, and decide on fall event
infringing on privacy, particularly in bedrooms and bathrooms. from the person’s proximity to the floor. The radar is ceiling-
Radar offers remote monitoring, without requiring the user to mounted. We present a technique for detecting breathing
carry any electronic device, and it also protects privacy better that can be integrated with the detection of other movements
than a camera. Researchers have looked into the potential for localization, including when the person is inert (such as
fallen on the floor). (ii) A study of human radar cross section
Manuscript received November 30, 2019; accepted December 30, (RCS) is presented along with new results on the statistics
2019. Date of publication January 17, 2020; date of current version
April 3, 2020. The associate editor coordinating the review of this article of human RCS including the case of circular polarization.
and approving it for publication was Dr. Ying Zhang. (Corresponding (iii), A neural network classifier is used to train on radar
author: Abhijit Bhattacharya.) Doppler signatures to differentiate people from other moving
Abhijit Bhattacharya is with Sierra Wireless Inc., BC V6V 3A4, Canada
(e-mail: [email protected]). entities present in an indoor environment. Using the receiver
Rodney Vaughan is with the Engineering Science Department, operating characteristic (ROC) metric, we show that applying
Simon Fraser University, Burnaby, BC V5A 1S6, Canada (e-mail: the classifier along with range tracking has the potential to
[email protected]).
Digital Object Identifier 10.1109/JSEN.2020.2967100 reduce false alarms and improve system reliability.

1558-1748 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 BHATTACHARYA AND VAUGHAN: DEEP LEARNING RADAR DESIGN FOR BREATHING AND FALL DETECTION 5073

The rest of the paper is as follows. Section II describes the where the geometric (ie., optical) cross section represents the
design and verification of the frequency modulated continuous cross-sectional area (A) of the target as viewed from the
wave (FMCW) radar used in this work. Section III presents radar antenna, and it models how much power the target
our method of detection for breathing and fall, along with will intercept from the incident power density Pi at the
experimental results. Section IV, discusses the neural network target. The reflectively refers to the fraction of the intercepted
architecture and presents radar signal classification results, and power (Pint ercept ed ) that is scattered by the target (Pscat t er ).
Section V concludes our present work. It depends on the material properties of the target at the
II. R ADAR S YSTEM D ESIGN & V ERIFICATION frequency of operation, the angle of incidence of the incident
wave and its polarization. The directivity is the ratio of power
The Radar Range Equation (RRE) in (1), is the starting
that is backscattered, Pbackscat t er , (for the monostatic radar
point for radar system design [17]:
case, which is of primary interest here) by the target to the
 1/4
Pavg (ηant Dant )2 (c/ f o )2 σav tint e−2 α Rmax power that would have been backscattered by a hypothetical
Rmax = (1) isotropic scatterer, and these power terms are expressed as
(4π)3 kTs F(S/N)1 L f L s
power per solid angle (Watts per steradian). Clearly, the tar-
where Rmax is the estimate of maximum detection range
get’s size, shape, and material properties all contribute to its
(in meters) for certain probability of detection and certain
RCS. Using these individual terms in (2),
false-alarm probability; Pavg is the rms transmit power of
CW radar in Watts; ηant denotes total antenna efficiency; Pscat t er Pbackscat t er
σ = A× × (3)
Dant is directivity of the antenna; f o is operating frequency Pint ercept ed (1/4π)(Pscat t er )
of the radar in Hertz; σav is target’s average radar cross Pscat t er Pbackscat t er
section in square-meters; tint is the integration time at the ∴σ = A× × (4)
(A)(Pi ) (1/4π)Pscat t er
radar’s Doppler filters in seconds; α is the coefficient of
atmospheric attenuation due to oxygen, water vapor etc.; and cancelling the like terms in (4) yields
k is Boltzmann’s constant; Ts denotes system temperature in Pbackscat t er
Kelvin; F is noise factor of the entire receive chain including σ = 4π
Pi
the antenna and other noise contributors; (S/N)1 is the signal- Backscatter per unit solid angle
to-noise power ratio per chirp required for detecting a sine- = 4π .
Power per unit area of incident wave at the target
wave (i.e., a non-fluctuating target) in additive white Gaussian (5)
noise; L f corrects for loss due to the target’s radar cross
section fluctuations; L s accounts for the other losses in the The power delivered to the surface area of a unit solid angle at
system such as loss due to imperfect IF-filtering, loss due a distance Rr from the target can be written as Ps Rr2 , where Ps
to the target straddling the Doppler frequency bins, antenna is the power per unit area in the backscattered wave arriving at
beam-pointing loss, and also field degradation of hardware the receiving antenna. Then, from (5), using the fundamental
components. relation of the average power density of a plane wave to
The RRE shows the link between the target characteristics, its electric field strength (E) and the intrinsic impedance
propagation medium, and the radar’s hardware and signal of the propagating medium (η) given by P = (1/2)(|E|2/η),
processing. It gives a glimpse into the radar’s performance we arrive at the familiar form of RCS expression, e.g, [17],
with even approximate (initial) values of its parameters, and
|E s |2
helps to provide intuition before getting overly concerned σ = 4π lim Rr2 (6)
about its precision. While we have little control over some of Rr →∞ |E i |2
the parameters such as the target characteristics, σav and L f , where the superscripts i and s refer to the transmitted fields,
other parameters, typically the transmit power (Pavg ), antenna meaning “incident” fields at the target and “scattered” fields at
directivity (Dant ), and operating frequency ( f o ), can be traded the radar’s (receive) antenna. The limit in the above equation
for achieving performance such as the detection range or cov- emphasizes a far-field assumption. The scattered fields, for
erage. Thus the approach adopted in this paper is to start with transmitting an arbitrary linear polarization, can be written in
the approximations (in F, L s , and even σav ), and then to refine terms of the polarization basis (φ and θ ) as
the design through iteration. Finally, some safety cushion (by
increasing the transmit power) is added to cover the statistical E φs = sφφ E φi + sφθ E θi
nature of some of the parameters that went into (1), to come E θs = sθφ E φi + sθθ E θi . (7)
up with a working design for the purpose of breathing and fall
detection in a real-world setup. This relationship, using the polarization scattering matrix, [s],
is
A. Radar Cross Section (RCS) of a Human  s     
Eφ   E φi sφφ sφθ E φi
A critical parameter in (1) is the target’s radar cross section = s = . (8)
(σ ). The RCS can be expressed in terms of some physical E θs E θi sθφ sθθ E θi
aspects of the target that determine its value ( [18], [19]), Substituting (8) in (6), the RCS can be expressed as
σ = Geometric Cross Section × Reflectivity × Directivity    
σφφ σφθ 2 |sφφ |
2 |sφθ |2
= 4π Rr (9)
(2) σθφ σθθ |sθφ |2 |sθθ |2

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 5074 IEEE SENSORS JOURNAL, VOL. 20, NO. 9, MAY 1, 2020

where the entries of the RCS matrix (e.g., σφφ and σθφ )
refer to transmitting one polarization (e.g., φ), and observing
the scattered fields in that or the orthogonal polarization
(i.e., φ or θ ).
RCS measurements need an elaborate test setup [20]. There
are some recent reports on RCS at higher microwave and
mm-wave frequencies which are of interest to automotive
applications ( [21], [22]), but there is scarce information
on for human radar cross section around our frequency of
interest at S-band [23], [24] Also, in practice, there is no
clear direction for choosing the values for σav and L f
in (1). [17] proposes an approach based on the classical
analysis by Swerling [25]. Its advantage is that once a target
class can be adequately represented by one of the Swer-
ling’s models, the associated probability density functions
(that have been found to correspond well to a variety of
real-world targets), have convenient tables of detection prob-
ability. Moreover, how the variations in cross section with
time affect the detection probability has been bracketed in
Swerling’s treatment of his models. A disadvantage in this
approach is that some knowledge of the statistical nature of
a target’s radar returns is required. Since Swerling’s work
of the 1940s, advances in numerical electromagnetic (EM) Fig. 1. (a) Comparison of complex reflection coefficients of planar
tools have made it possible to get accurate estimates of RCS multilayer model of human torso with equivalent single dielectric model,
for electrically large objects. This requires modest computing (b) monostatic RCS of torso in azimuth for different dielectric models. σVV
(≡ σθθ in azimuth) denotes the RCS when both incident and scattered
resources - commodity computers completing the task within fields in (9) are vertically polarized, whereas σHV (≡ σφθ in azimuth)
a reasonable time [26]. Therefore, an in-depth analysis of RCS is the case when the incident field is vertical and the scattered field is
of human based on numerical simulations was undertaken to horizontally polarized.
facilitate the radar design. For these numerical experiments,
we use WIPL-D [27] which is a method-of-moments based where eq n and n
mult are respectively the complex reflection
surface integral equation solver. WIPL-D uses: quadrilateral coefficients of the equivalent dielectric and the multilayer chest
mesh elements and higher order basis functions (HOBF) that models for angle of incidence n (in degrees) and N denotes
reduce the number of unknown coefficients; and techniques the total number of angles.
like “shadow plane”; and local mesh sizing. It also allows Fig. 1(a) compares the reflection coefficients of the equiv-
CPU, GPU and GPU cluster accelerations, and so electri- alent dielectric (r = 118.5 and σc = 2.5Siemens / m) with
cally large structures can be solved in a relatively short the multilayered model for all angles of incidence and for
time [28]. both transverse electric (TE) and transverse magnetic (TM)
The human body is made of several layers of cells and polarization of the wave. Although the phases of the reflection
tissues [29], [30]. At higher microwave frequencies (X-band coefficients of the two models deviate around the pseudo-
and above) EM-waves do not propagate much further than Brewster angle (for TM polarization), there is still reasonable
the outer (skin depth) layer, making it possible to use a agreement. It was found that the 3D correlation [32] of
simpler “homogeneous skin model” for full-sized human body RCS from these models is well over 99%, confirming a
to simplify numerical computations [31]. At lower frequencies good match. Fig. 1(b) shows monostatic (backscatter) RCS
this simplification will not hold, with the skin, fat, and muscle in azimuth for a human torso for both models.
playing a role in the radar return. A complex multi-layered Fig. 2(a) shows monostatic RCS of a full-sized human body
model is hard to build and takes much longer to simulate, model with the equivalent dielectric derived above, for the
therefore we replace the multilayered structure with a single azimuth cut. Here, V and H denote θ - and φ- components
equivalent homogeneous dielectric (r and conductivity σc ). of the fields respectively. The co-polar RCS (σV V and σ H H )
The approach is to minimize the rms of the difference (RMSE) have similar values as the model is electrically large and com-
between the the complex reflection coefficients of a planar plex in S-band frequencies. This observation agrees with the
multilayer version of the human torso and the equivalent single experimental results of [23]. Furthermore, it is interesting that
dielectric model, as follows: the cross-polar RCS (σ H V and σV H ) are the same which can
be explained from the reciprocity theorem for linear passive
RM
SE medium [33]. Fig. 2(b) hints how the RCS response could

1 90 average out during a chirp (covers a band of frequencies),
= | eq,T E − mult,T E | +| eq,T M − mult,T M |
n n 2 n n 2 filling out the deep nulls thus improving target’s visibility.
N Fig. 3 shows cumulative distribution function (CDF) of
n=0
(10) co-polar RCS for a family of human body models

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 BHATTACHARYA AND VAUGHAN: DEEP LEARNING RADAR DESIGN FOR BREATHING AND FALL DETECTION 5075

Fig. 4. PDF of co-polar RCS for full-scale human body model with
Fig. 2. (a) Monostatic RCS of a full-scale human body model with equivalent dielectric (a) overall distribution, (b) distribution near azimuth.
equivalent dielectric at 2.45 GHz, (b) Co-polar RCS variation within
radar’s bandwidth.
This means that most of the backscatter can be explained
by Swerling case 1, which assumes a target with a “few”
independent scatterers of comparable returns, i.e., of similar
size.(NB., the exponential distribution for power corresponds
to the Rayleigh distribution for magnitude, [34], so Swerling’s
“few” corresponds to “many”) Moreover, the remaining 10%
to 15% of the backscatter that is primarily from the front and
back of the body close to the azimuth plane (with a standard
deviation of about 15o in elevation) conforms well with
Swerling case 3 which assumes one dominant scatterer which
is physically larger than the others together with few weaker
ones (NB., the Rice distribution for magnitude). Therefore,
based on the application, one should be able to choose the
right model for human RCS distribution. For example, when
Fig. 3. CDF of RCS for a family of human body models with the equivalent
dielectric. verifying radar’s range performance (in subsection D) we
chose the model in fig. 4(b) (Swerling case 3). But considering
(labeled alphabetically) for all directions in φ and θ . This the application to fall detection where the antenna is mounted
study was taken up to assess the variability of human RCS with on the ceiling of a room, distribution in fig. 4(a) (i.e., Swerling
height (190 cm for the tall man to 120 cm for the child), body case 1) is more appropriate.
shape (slim or fat) and posture (sitting, standing, extending The monostatic RCS described above is the far-field case
arms etc.). Interestingly we see that for about 90% of the where the target is in the far-field of the radar and vice versa,
angles, the RCS stays within about 0.5 square meters, with so it assumes that the target is illuminated with a uniform plane
the exception of the tallman. For the tallman extending wave. But in our application the antenna is mounted on the
his arms we notice increased response to the φ polarization. ceiling not far from the human body, so the target may be in the
Fig. 3 also indicates the likely need for higher transmit power near-field of the radar, and illuminated with a spherical wave-
for detecting breathing which of course is restricted to small front that has also has amplitude taper from the representing
movements of the torso. beam shapes. The potential impact of the near-field case is
Fig. 4(a) shows the distribution of co-polar RCS for the investigated in Fig. 5 which shows RCS distribution in the
model in Fig. 2. It can be seen that an exponential power dis- near-field, from a simulation using WIPL-D of a numerical
tribution represents this RCS well for about 90% of the angles. model of the linearly polarized antenna used in our radar.

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 5076 IEEE SENSORS JOURNAL, VOL. 20, NO. 9, MAY 1, 2020

Fig. 6. RCS comparison for linear and circularly polarized incident waves
(a) in azimuth plane (θ = 90o ), (b) in elevation plane(φ = 0o ).
Fig. 5. PDF of co-polar near-field RCS for full-scale human body
model with equivalent dielectric. This does not include the contributions
corresponding from the near-azimuth directions, which are off the scale.

It shows a similar RCS distribution level to that of Fig. 4(a)

on average (σav = 0.1031 square meters), so the assumptions
of the far field seem reasonable.
For a given frequency and given orientation of the radar and
the target, the polarization scattering matrix [s] of (8) contains
the scattering properties of the target. A scattering matrix is
established by any polarization basis (e.g, orthogonal linear
polarizations φ and θ ). It can be transformed to any other basis.
For example, the circular scattering matrix elements from the
linear scattering elements are given by ( [35])
    
sR R sR L   sφφ sφθ 1 0  −1
= T T (11)
sL R sL L sθφ sθθ 0 −1 Fig. 7. (a) Block diagram of the FMCW radar, (b) receiver noise
figure calculations.
where
      
  1 1 −j E sR s sR L E iR
T = √ and = RR ,
2 1 j E sL sL R sL L E iL
and subscripts R and L denote right-hand and left-hand circular
polarizations of the waves respectively. Fig. 6 compares the
backscatter from a human body for this example of circularly
and linearly polarized waves. In the azimuth direction, where
the RCS response is from a quasi-planar surface, the pair of
co-polarized linear responses (V -to-V and H-to-H) and the
pair of cross-polarized circular (R-to-L and L-to-R) responses
are each similar, as expected from a dominant single reflection
from a reasonably conductive surface. So an antenna system
with proper sense of circular polarization in transmit and
receive (or any other polarization basis) could replace the
linearly polarized antenna in our experiments, and although the Fig. 8. The radar system prototype.
RCS responses will differ between polarization bases, the radar
information is the same. commercial radar development kits for reasonable price,
e.g., [40]–[42].
For this work we designed the sensor from scratch. Fig. 7(a)
B. Hardware Design Considerations shows its high-level block diagram, and Fig. 8 shows our
There are plenty of texts available that describe various radar prototype. We use a single antenna for transmit and
aspects of radar theory, e.g., [17], [35], [36]. The focus has receive a configuration with scant coverage in the literature.
been mostly been on pulse radars, and there are very few Using single antenna puts some serious constraints on antenna
books devoted to FMCW techniques, cf., [37], [38]. Of these, return loss (20 dB or better), to keep transmit leakage from
[37] is better for hands-on design but lacks the necessary saturating the receiver. This makes the antenna design chal-
theoretical basis, specifics of many of the necessary design lenging, especially when antenna volume is restricted and
decisions, and verification techniques which are crucial in a large frequency bandwidth (BW ) is needed to improve
developing an intuition into the design process. A popular range resolution (inversely proportional to BW ) at the same
resource is [39] and there are some companies offering time. But the benefits, including a reduced antenna space

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 BHATTACHARYA AND VAUGHAN: DEEP LEARNING RADAR DESIGN FOR BREATHING AND FALL DETECTION 5077

Fig. 9. Signal processing flow of FMCW radar.

(especially for arrays), lower hardware cost, easier mounting, C. Signal Processing for the Radar
easier beam-pointing to the target, justify further investiga- The ground work for FMCW radar range-Doppler
tion. Using a higher operating frequency is typically more processing in Barrick’s seminal report from 1973 [48]. Later,
expensive, but also allows a lower relative bandwidth and Stove [45] presented the moving target indicator (MTI)
physically smaller antenna. Another approach is a shared processing for FMCW signals. The following is a condensed
aperture, simultaneous transmit and receive antenna system, description of those concepts in context of our radar.
e.g., [43] but this requires extremely high isolation, which is Referring to Fig. 9, the local oscillator signal (LO), and the
very difficult over a reasonable bandwidth. delayed and attenuated received signal (Rx) from a target at
Another area that is not covered well in the literature range R during the n t h chirp can be expressed as
is the calculation of the overall noise factor (F), an
essential entry in the RRE. A receiver chain has several s Ln O (t f , t) ≈ A sin 2π f L (nT + t f − t L O )
distinct sections − the radio frequency (RF) part, intermediate-
frequency/baseband (IF/BB), and analog-to-digital converter K (t f − t L O )2
+ + φ0 (12)
(ADC), and the digital signal processing (DSP). To perform 2
a full signal-chain noise analysis requires conversion between B
the various metrics conventionally used for the different parts s nRx (t f , t) ≈ sin 2π f L (nT + t f − tT R − td )
R2
of the chain - RF components uses the noise figure (N F,
K (t f − tT R − td )2
in dB) to specify a noise contribution, the BB component + + φ0 (13)
noise is typically 2
√ specified using a magnitude noise spectral
density (nV / H z), and ADC noise is primarily specified as where A and B represent amplitude of the signals in volts;
an equivalent SNR (S N R, in dB). Moreover, the common (nT + t f ) is the time (t) including the so-called slow-time
assumptions of a constant impedance and proper source/load (clock time) and the fast-time (propagation time), t f , with
termination, seldom applies in practice. Fig. 7(b) shows the 0 ≤ t f ≤ T ; f L is the transmitter frequency at t = 0;
N F calculations for our design [44]. Our homodyne receiver K = 2BW/T is the slew rate; td is the two-way propagation
has a relatively high N F which comes from flicker noise delay; and φ0 is initial phase in rad/s.
(1/ f noise) dominant in amplifiers and filters with low IF The beat signal at node 1, after the mixer and filtering
frequencies. A superheterodyne receiver improves this noise (including the anti-aliasing filter), neglecting the terms with
contribution at the cost of increased complexity. We have tT2 R , t L2 O tT R td , and td2 , is
not investigated the effects of non-idealities in the transmit-
AB
ter on our system performance, which may bring impair- sbn (t f , t) ≈ cos 2π( f L + K t f )(td + tres ) (14)
ments especially for multi-target detection. For example, treat- 2R 2
ments of non-linearity in FMCW transmitters can be found where tres = tT R − t L O is the residual delay in the hardware.
in [45]–[47]. Substituting td = 2 R + x(t) /c = 2 R + V (nT + t f ) /c

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 5078 IEEE SENSORS JOURNAL, VOL. 20, NO. 9, MAY 1, 2020

in (14), and after some manipulations,

AB 2K R
sbn (t f , t) ≈ cos 2π( + K tres + f D )t f
2R 2 c
4π f L R
+ 2π f D nT + 2π f L tres + (15)
c
where x(t) is the displacement due to target’s velocity (V ), c
is speed of light in vacuum in m / s, and f D = 2 f L V / c is
the Doppler frequency. Next, for MTI processing, the returns
from n t h and (n − D)t h chirps are subtracted as shown in
Fig. 9. Applying trigonometric identities and noting that the
terms 2 π f L tres and 4 π f L R / c add a constant phase (φ1 ),
Fig. 10. Frequency response of MTI filters (see Fig. 9) for breathing
the output at node 2, is and movement detection. For example, the red line (top left of the
sb(n) (t f , t) − sb(n−D) (t f , t)
“movements” lines) represents the response of a two-chirp canceller with
D = 3.
∝ sin π D f D T
2K R
× sin 2π + K tres + f D t f +(2n − D)π f D T +φ1
c
(16)
where it is assumed that there are no sweep-to-sweep fluc-
tuations in the amplitudes of the signals (Swerling 1 or
3 target) during the MTI processing interval, and, as noted
above, no non-linear effects within the sweep. Furthermore,
the changes in target’s range and velocity are also negligible in
this interval. The sin(π D f D T ) term in (16) is the frequency
response of the MTI filter. The received signal is a super-
position of chirps coming with different delays from different
objects in the radar’s view (plus any transmit leakage). In such
cases, the MTI output peaks at the Doppler frequency of
the desired target ( f D = 1 / 2 D T ), and the returns from
other objects (e.g., static clutter in a room from floor, walls,
furniture etc.) are attenuated. Fig. 10 shows the MTI filters
used in this work. The frequency response of these filters were
set through physical experiments.
Signals in the form of (16) for the consecutive chirps
are arranged in columns of a matrix (at node 2). Then
a Fast Fourier Transform (FFT) of the n-th column
will peak at ± 2 K R / c + K tres + f D with associated
phases ± (2 n − D) π f D T + φ1 . Considering only positive
Fig. 11. Experimental results of Doppler and range verification against
frequencies, this can simply be written as a simple pendulum.

e+ j (2n−D)π f D T +φ1
δ(ω − ω0 ) (17)
D. Calibration and Verification of Radar Performance
where ω0 = 2π 2 K R / c + K tres + f D gives the range.
(Ignoring errors due to Doppler contributions, and correcting In this section we investigate how an experimental range
for any internal delays, tres , the range is available from estimate from our radar compares to the prediction from our
this angular frequency.) With a velocity of the human torso pen-and-paper design. First, the sensor requires calibration and
1.5 m / s, the error in the estimation of R due to the contribution verification of its range and Doppler outputs. The additional
from the Doppler term will be ∼ 0.04 m for our experimental path length corresponding to the residual delay tres in (17) was
setup in II-D. Interestingly, from (17) we see that along estimated by simply putting a short-circuit at the antenna con-
the rows of the matrix (i.e., in slow-time) there are phase nector (which zeros the range, ie., R = 0), and taken out from
increments (= 2 π f D T ), proportional to the target’s Doppler further range calculations. A range estimate was then verified
frequency, between the successive column entries of the matrix using an RF cable of known electrical length and velocity.
at node 3. Thereby, a second FFT along the rows (for a number Fig. 11 shows Doppler and range plots of a simple test pendu-
of sweeps N, whose total duration is called the coherent lum in motion using the radar. The close agreement between
processing interval (CPI)) gives the Doppler frequency, f D these experimental results and the theoretical estimates con-
Hz, of the target (node 4 of Fig. 9). Of course, for a better firms that the hardware and the signal processing is working as
Doppler resolution (= λ/2N T ) a longer CPI is required. desired.

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 BHATTACHARYA AND VAUGHAN: DEEP LEARNING RADAR DESIGN FOR BREATHING AND FALL DETECTION 5079

assumed to be 3dB simply following standard practice. The

relative dielectric constant of the ground was taken from [51].
Similarly, other parameters were set based on the targets’
physical and Doppler characteristics. The sweep time was
set to (T) = 6 ms, the frequency bandwidth was (BW) =
250 MHz centered at 2.425 GHz, resulting in a theoretical
range resolution of about 0.6 m and a maximum unambiguous
velocity of about 5 m / s which we found sufficient to cap-
ture the human movements in our experiments. Experimental
results (in the inset of Fig.12), when transmitting at −10 dBm,
indicate that the radar is in reasonable agreement with the
designed performance. Finally, in light of the data in Fig. 4(a)
Fig. 12. Theoretical estimates of range performance of the radar vs. and considering additional losses in the antenna RF cables to
measured results (inset) from the field test. The radar antenna was be used for test mounting, the transmit power was raised to
oriented for vertical polarization. 0 dBm for experiments with breathing and fall detection.
The RRE (1) was formulated with airborne targets in mind,
III. B REATHING AND FALL D ETECTION
and does not take into account the ground reflections present in
our field test. (see the inset of Fig. 12 for the field test setup.) The results in Fig. 12 demonstrated that the radar is capable
Therefore, a model was developed with ground reflections, of sensing human movement (walking) and the target range.
for our range calculations. Adapting the usual 2-ray model, In this section we present a technique for detecting breathing
e.g., [34], [49], the power at the target can be written as that is suitable for real-time implementation and can be readily
 2  √ 2 integrated into the range processing to detect falls in indoor
λ  G l T E/T M  j φ
 situations when the subject is static.
Pt arget = Pavg  + Gr e  (18)
4π Rd R We start with the general idea of breathing detection from
  i 
term1 current literature. If we substitute td = 2 R + x (t) / c in
(5), where x (t) represents periodic displacement of human
where Pavg is the time-averaged transmit power of the radar,
chest / torso due to breathing activity, (and ignoring tint for
λ is the free-space wavelength, G l = G t d (4 π σav / λ2 ),
time being) we get
G r = G t i (4 π σav / λ2 ), and G t d is transmit antenna total
gain in the direct path, G t i is transmit antenna total gain fast-time
  
towards reflected / indirect path, T E / T M is the reflection n AB 2K R 2K x (t)
coefficient of the ground for TE or TM wave [50] φ is the sm (t f , t) ≈ cos 2π tf + tf
2R 2 c c
phase difference between the direct path (Rd ) and the reflected 2 fL 2 fL R
path (Ri ). + x (t) + . (21)
Applying reciprocity, the backscattered power at the radar  c   c
is slow-time
 From (9) the phase of the signal in slow-time can be written
 2  G 2
λ  l T E/T M  
j φ as
Pscat t er = Pt arget  + G r e  (19)
4π Rd Ri 4π f L 4π f L
   = x (t) + φ2 = ab cos(2π f b t) + φ2 (22)
term2 c c
where ab and fb are respectively the amplitude and frequency
where G l = G rd , G r = G ri are respectively the receive
of breathing. In practice, there is also a signal from the
antenna total gain in the direct and reflected path directions.
heartbeat, so () is superposition of two quasi-periodic sig-
The SNR at the radar can be written as
nals. In the frequency domain, the problem becomes finding
S Pavg × ter m1 × ter m2 × tint
= . (20) the fundamental frequency from a mixture (of harmonics
N kTs F L s and intermodulation products) of periodic signals [52], [53].
where tint , k, Ts , F and L s were defined earlier for (1), and Moreover, from (22), the amplitude of the phase variation is
ter m1 and ter m2 are in equations (18) and (19). Considering proportional to the frequency f L , which is why most radars
the RCS characteristics at the frequencies of our interest, for vital-sign detection, use mm-wave frequencies rather than
the above formulation assumes a constant (=average) radar lower microwave frequencies [54].
cross section for the target along the direct and reflected paths. Complex signal processing and a high operating-frequency
Fig. 12 shows the theoretical estimates of the radar range both affect the real-time capability and detection range of
performance from (20), and compares with that of (1). In these present techniques. But here we demonstrate breathing detec-
calculations we used the antenna gains taken from our own tion directly from range processing. As shown in Fig. 13(a),
pattern measurements (with a peak gain of 6 dBi). The average breathing leaves a signature on the range heat-map. As the
RCS (σav ) was taken to be 0.7 m 2 according to our simulation chest undergoes inhalation and exhalation, the Doppler fre-
results of Fig. 4(b). The noise factor (F) of the receiver was quency of the radar echoes moves in and out of the MTI
taken from Fig. 7(b). Other losses in the system (L s ) were filter’s response, thus modulating the intensity of the returns

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 5080 IEEE SENSORS JOURNAL, VOL. 20, NO. 9, MAY 1, 2020

Fig. 14. RTI plots and breathing estimates. Different MTI filter detects
different breathing rates.

Fig. 13. (a) RTI plot of breathing, (b) FFT of the smoothed trace possible to cover bigger rooms than that of the coverage of
(breathing frequency was derived from it).
the experiments demonstrated here.

and creating periodic fluctuations. Breathing frequency can be

IV. A DDING I NTELLIGENCE TO THE R ADAR
estimated by tracing these intensity fluctuations, and then from
the smoothed quasi-periodic trace, picking the frequency with Beside target detection, radar systems often need to distin-
highest amplitude through an FFT, as shown in Fig. 13(b). guish between targets. For example, a classical air surveillance
Different breathing rates can conveniently be detected with radar is required to tell the difference between a passenger air-
reasonable accuracy by choosing appropriate filters for the craft and a fighter jet before taking further action [55]. Radar
MTI filter bank. Fig. 14 shows few more examples where the on an autonomous vehicle should be a able to differentiate
subjects were facing the main beam of the radar at a distance between various objects in its path to navigate safely [56].
of about 2.5 m. Finally, breathing and movement detection are Automatic recognition of birds and drones can reduce operator
combined to identify falls. workload in an anti-drone radar [57]. Similarly, for our case of
A fall is a complex physiological phenomena, and the exact indoor fall detection it would be helpful to tell apart a person
way someone falls varies widely based on the circumstances. from other moving entities in the room, and track accordingly.
But one aspect that nearly all falls have in common, fol- This can reduce false-positives and increase system reliability.
lows the observation that “people fall to the floor”. Hence, Traditionally, recognition is done in three steps. The first
by mounting the radar on the ceiling (e.g., of a washroom), task is to define target class. Recognition problems can be
and by monitoring the range we should be able to tell if two-class or multi-class. The next step is to choose target
someone has fallen, simply from their proximity to the floor. features that enhance class separability. This is commonly
Fig. 15 shows one such arrangement. The antenna (in inset) is referred as “feature engineering” and usually requires con-
mounted on the ceiling of a small (washroom sized) area, and siderable knowledge of the problem domain. The final step is
illuminating the area with a single static beam. The person making a decision regarding the class of the target. Regardless
is standing right under the antenna beam and performing a of the approach, be it deterministic (some generalized deci-
succession of normal activities such as picking up object from sion function) or statistical (e.g., Bayes classifier), a suitable
the floor, squatting and so on. The radar is able to tell the decision criterion is required. Recent advances in machine
distance of the person with sufficient accuracy throughout the learning (especially deep neural networks) have opened a new
range of her motions. Furthermore, the radar is capable of paradigm where the system is presented with many data-
detecting the presence of the person and the range accurately, points (training examples) relevant to a task, and it determines
solely from the breathing when subject is lying otherwise the statistical rules to automate the task on new (never-seen-
motionless on the floor. It is reasonable to expect that using before) data [58]. This elimination of some critical steps in the
automatic tracking based on the continuity of human motion traditional approach, means that deep neural networks have
and practical constraints on speed in indoor settings, it will be augmented classical signal processing significantly. The areas

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 BHATTACHARYA AND VAUGHAN: DEEP LEARNING RADAR DESIGN FOR BREATHING AND FALL DETECTION 5081

Fig. 15. Experimental fall.

impacted to date has been pattern recognition, such as visual

recognition, medical imaging, pedestrian detection, speech and
music analysis, et. al. The following study is a form of human
recognition in indoor scenes using radar echoes and a deep
neural network model.
A. RadarNet: Convolutional Neural Network for Radar
Target Classification
The successful application of deep neural networks, partic-
ularly convolutional neural networks (CNN), have prompted
their use in diverse areas of research. In 2015, Piczak [59]
presented results of applying a CNN to environmental sound
classification (ESC) problem. Remarkably, results of his sim-
ple and shallow network were comparable to state-of-the-art
approaches of the day. Many researchers have now improved
upon the results mainly by using deeper and more compli-
cated networks sequentially extracting more feature maps [60].
Recently, [61] reported ∼98% classification accuracy on the
same dataset using a relatively small but wide neural network
by extracting more feature maps at each level. They associate
this significant increase in accuracy (from 73% in the original
work [59]) primarily to the network’s ability to extract infor-
mation more efficiently than sequential architectures. All these
CNN models take spectrograms of environmental sound clips
as input. Interestingly, these spectrograms resemble Doppler-
time-intensity (DTI) plots of our radar returns, prompting a
look at the problem of classifying people and other moving Fig. 16. The RadarNet architecture.
entities in a radar scene using a CNN.
To keep things simple, the task is framed as a two-class binary cross entropy, a popular cost function for two-class
recognition problem with “person” being one class and “pet” classification problems. Inference was drawn directly from the
being the other. Moreover, in “pet” we looked at the radar sig- final (average-) pooling layer. We did not use any densely-
natures of dogs only. Fig. 16 shows our CNN model, which for connected layers, which helped to keep the model size small.
brevity we refer to simply as “RadarNet”. RadarNet consists of This is particularly important when considering implementa-
one convolutional block followed by three inception modules tion with limited hardware resources, especially in embedded
[62] stacked together with pooling layers in-between. We used applications. Moreover, during training, about half of all output

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 5082 IEEE SENSORS JOURNAL, VOL. 20, NO. 9, MAY 1, 2020

Fig. 18. Spectrograms of radar returns indicating differences between

Fig. 17. Spectrogram data slicing strategy.
radar signature of (a) human and (b) dog.

features were dropped out of all three inception modules and

from the final average-pooling layer. This leveraged sparsity
of the convolutional kernels, and using dropout regularization
[63], overfitting was tackled in the training using a limited
number of examples.
As mentioned, a key advantage of deep learning is that
it can find useful features in the training data without any
need for manual feature engineering. But this can only be
achieved when large numbers - hundreds to thousands - of
training examples are available (the supervised learning).
For radar signal classification there is no public dataset,
unlike ESC (UrbandSound8K dataset) and computer vision
(ImageNet dataset), etc. Therefore, a measurement campaign
was undertaken to record radar Doppler signatures of indi-
viduals and dogs from various angles in indoor environment
[64], [65]. Fig. 17(a) shows a sample DTI plot of a person
walking towards the radar and then away from it, periodically.
Fig. 17(b) shows the same for a trained dog. After studying
the Doppler and micro-Doppler variations across the dataset,
we decided to slice up the experimental data (like those in
fig. 17) into durations of T p = 2 s to feed the CNN. Then,
the lag (δ) between those portions was set based on the
image similarity [66] measures of fig. 17. Finally, these image
segments (discarding the silent segments) were resized to Fig. 19. (a) Accuracy and loss figures of RadarNet for hold-out validation,
and (b) ROC curve for the same. A large area under the curve (AUC)
56 pixels × 56 pixels and collected sequentially under cor- indicates good classification performance.
responding class labels, amounting to 2000 samples in each
class (balanced classes). Fig. 18 shows some examples from apart as the test set, and the training was on the remaining
each class. data. This way the overlap between training and test sets
For training and testing / validation, two schemes were was kept to a minimum (given both sets represent the overall
adopted − hold-out validation and 10-fold cross-validation. data well). Fig. 19(a) shows the validation accuracy (∼95%)
In hold-out validation, 30% of the data in each class were set obtained from RadarNet. Next, in 10-fold cross-validation,

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 BHATTACHARYA AND VAUGHAN: DEEP LEARNING RADAR DESIGN FOR BREATHING AND FALL DETECTION 5083

the data was shuffled and split into 10 equal partitions. One WIPL-D simulations. They also gratefully acknowledge the
of these partitions was set apart for evaluation, and training reviewers’ comments and suggestions.
was on the remaining data. Repeating this process for all the
partitions gave 10 different scores. An average accuracy of
∼98% with small variance aver the various train-test splits was R EFERENCES
observed. The other metric used (Fig. 19(b)) was the receiver [1] J. Dickmann et al., “‘Automotive radar the key technology for
operating characteristic (ROC), a classical technique first used autonomous driving: From detection and ranging to environmental
in world war II for radar signal analysis. Our ROC was understanding,”’ in Proc. IEEE Radar Conf. (RadarConf). May 2016,
pp. 1–6.
obtained by plotting probability of detection (probability of [2] M. Jian, Z. Lu, and V. C. Chen, “Drone detection and tracking based
detecting a “person” when present) against probability of false on phase-interferometric Doppler radar,” in Proc. IEEE Radar Conf.
alarm (probability of declaring “person” when it is “pet”) at (RadarConf), Apr. 2018, pp. 1146–1149.
[3] C. Gu, J. A. Rice, and C. Li, “A wireless smart sensor network based
various thresholds. The ROC curve shows our CNN classifier on multi-function interferometric radar sensors for structural health
has a very low false positive rate, desirable for applications monitoring,” in Proc. IEEE Top. Conf. Wireless Sensors Sensor Netw.,
such as fall detection or other health monitoring. Jan. 2012, pp. 33–36.
[4] D. A. Robertson et al., “A high frame rate, 340 GHz 3D imaging
High classification accuracy confirms that the radar spec- radar for security,” in Proc. IEEE Radar Conf. (RadarConf), Apr. 2018,
trograms are sufficiently informative to tell apart a person pp. 0055–0060.
from a dog (and potentially from other moving objects). [5] S. Skaria, A. Al-Hourani, M. Lech, and R. J. Evans, “Hand-gesture
recognition using two-antenna doppler radar with deep convolutional
Our data samples of persons and dogs by no means cap- neural networks,” IEEE Sensors J., vol. 19, no. 8, pp. 3041–3048,
ture the entire spectrum of possible movements. The simple Apr. 2019.
dataset (capturing only walking, bending, and other large [6] Z. Zhang, Z. Tian, and M. Zhou, “Latern: Dynamic continuous hand
gesture recognition using FMCW radar sensor,” IEEE Sensors J., vol. 18,
limb movements of people) can be the other reason for the no. 8, pp. 3278–3289, Apr. 2018.
high scores in our experiments. In this study we did not try [7] C. Li, V. M. Lubecke, O. Boric-Lubecke, and J. Lin, “A review on recent
extensive hyper-parameter tuning (e.g., grid search) and model advances in doppler radar sensors for noncontact healthcare monitoring,”
ensembling which can be helpful for more complex multi- IEEE Trans. Microw. Theory Techn., vol. 61, no. 5, pp. 2046–2060,
May 2013.
class classification. Studying the effects of different values [8] R. Igual, C. Medrano, and I. Plaza, “Challenges, issues and trends in
of T p on classification accuracy can be helpful for real- fall detection systems,” Biomed. Eng. online, vol. 12, no. 1, p. 66, 2013.
time implementation, and has also not been experimentally [9] F. Adib, Z. Kabelac, D. Katabi, and R. C. Miller, “3D tracking via
body radio reflections,” in Proc. 11th USENIX Symp. Netw. Syst. Design
investigated. Moreover, we expect that analyzing range profile Implement. (NSDI), 2014, pp. 317–329.
along with the Doppler spectrum can enhance recognition [10] S. C. Mukhopadhyay, “Wearable sensors for human activity monitoring:
performance of neural network. These are subjects of future A review,” IEEE Sensors J., vol. 15, no. 3, pp. 1321–1330, Mar. 2015.
[11] C. Rougier, J. Meunier, A. St-Arnaud, and J. Rousseau, “Robust video
research. surveillance for fall detection based on human shape deformation,”
IEEE Trans. Circuits Syst. Video Technol., vol. 21, no. 5, pp. 611–622,
V. S UMMARY AND C ONCLUSION May 2011.
[12] J.-H. Lee, Y.-T. Im, and S.-O. Park, “Doppler radar sensing system of
We have presented the basic design of a radar and associated respiration and heart rate,” in Proc. 8th Int. Symp. Antennas, Propag.
signal processing, for fall detection. A motivating applica- EM Theory, Nov. 2008, pp. 706–708.
tion is the detection, classification and tracking for falls of [13] F. Adib, H. Mao, Z. Kabelac, D. Katabi, and R. C. Miller, “Smart homes
that monitor breathing and heart rate,” in Proc. 33rd Annu. ACM Conf.
elderly people living alone or away from immediate direct Hum. Factors Comput. Syst., 2015, pp. 837–846.
assistance. The radar is simple, low-cost FMCW, operating [14] B. Jokanovic, M. Amin, and F. Ahmad, “Radar fall motion detec-
at low microwave frequencies, viz., sub-6 GHz. The radar tion using deep learning,” in Proc. IEEE Radar Conf. (RadarConf),
sensor, with its deep-learning signal processing, was built and May 2016, pp. 1–6.
[15] Y. Tian, G.-H. Lee, H. He, C.-Y. Hsu, and D. Katabi, “RF-based fall
deployed, and the methodology of the testing was presented. monitoring using convolutional neural networks,” Proc. ACM Interact.,
We report the differentiation of people from their environment Mobile, Wearable Ubiquitous Technol., vol. 2, no. 3, p. 137, 2018.
using a neural network classifier, and locate their position with [16] B. Erol and M. Amin, “Effects of range spread and aspect angle on
radar fall detection,” in Proc. IEEE Sensor Array Multichannel Signal
good accuracy - a few cm for location, and much smaller Process. Workshop (SAM), Jul. 2016, pp. 1–5.
distances for breathing, corresponding to small fractions of [17] I. S. Merrill, Introduction to Radar Systems, vol. 7, 3rd ed. New York,
a wavelength. We have identified a few ways to make our NY, USA: McGraw-Hill, 2001, Ch. 1–6.
[18] G. W. Stimson, Introduction to Airborne Radar. Raleigh, NC, USA:
present system even more robust, and these are included for SciTech: 1998.
future development. Our demonstration sensor can be used [19] E. F. Knott, J. F. Schaeffer, and M. T. Tulley, Radar Cross Section,
not only to detect fall, but also for activity monitoring, for 2nd ed. Raleigh, NC, USA: Scitech 2004, ch. 3, pp. 63–74.
[20] J. Fordham et al., “RCS measurements in a compact range,” Microw.
example for in-home patient care, or even in home security Product Dig., Englewood Cliffs, NJ, USA, Tech. Rep., 2005. [Online].
and safety applications. With the rapid development of the Available: https://www.nsi-mi.com/images/Technical_Papers/Articles/
internet-of-things infrastructure, this demonstration contributes NSI-MI%20-%20RCS%20Measurements%20in%20a%20Compact%20
Range.pdf
to realizing the vision of a smart home that can unintrusively
[21] M. Chen and C.-C. Chen, “RCS patterns of pedestrians at 76-77 GHz,”
monitor our well-being. IEEE Antennas Propag. Mag., vol. 56, no. 4, pp. 252–263, Aug. 2014.
[22] N. Yamada, Y. Tanaka, and K. Nishikawa, “Radar cross section for
ACKNOWLEDGMENT pedestrian in 76 GHz band,” in Proc. Eur. Microw. Conf., vol. 2,
Oct. 2005, p. 1018.
The authors would like to thank Mr. T. Milosevic and [23] F. Schultz, R. Burgener, and S. King, “Measurement of the radar cross
Mr. B. Mrdakovic for their help and support with the section of a man,” Proc. IRE, vol. 46, no. 2, pp. 476–481, Feb. 1958.

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 5084 IEEE SENSORS JOURNAL, VOL. 20, NO. 9, MAY 1, 2020

[24] C. Le and T. Dogaru, “Numerical modeling of the airborne radar [51] W. C. Lee, Mobile Communications Engineering. New York, NY, USA:
signature of dismount personnel in the UHF-, L-, Ku-, and Ka-bands,” McGraw-Hill, 1982.
Army Res. Lab. Adelphi Md Sensors Electron Devices Directorate, [52] M. He, Y. Nian, and Y. Gong, “Novel signal processing method for vital
Adelphi, MD, USA, Tech. Rep. ARL-TR-4336, 2007. [Online]. Avail- sign monitoring using FMCW radar,” Biomed. Signal Process. Control,
able: https://www.arl.army.mil/arlreports/2007/ARL-TR-4336.pdf vol. 33, pp. 335–345, Mar. 2017.
[25] P. Swerling, “Probability of detection for fluctuating targets,” IEEE [53] C. Li and J. Lin, “Random body movement cancellation in doppler radar
Trans. Inf. Theory, vol. 6, no. 2, pp. 269–308, Apr. 1960. vital sign detection,” IEEE Trans. Microw. Theory Techn., vol. 56, no. 12,
[26] D. Lee, S.-H. Chi, H.-K. Jung, and D.-Y. Jang, “Comparison of com- pp. 3143–3152, Dec. 2008.
mercial simulation performance for efficient RCS analysis,” in Proc. Int. [54] T. Sakamoto et al., “Measurement of instantaneous heart rate using
Appl. Comput. Electromagn. Soc. Symp. (ACES), Mar. 2018, pp. 1–2. radar echoes from the human head,” Electron. Lett., vol. 54, no. 14,
[27] WIPL-D. Electromagnetic Simulation and EM Modeling Software. pp. 864–866, Jul. 2018.
Accessed: Jan. 15, 2019. [Online]. Available: https://wipl-d.com/ [55] Y. S. Zinoviev, A. Y. Pasmurov, D. B. Kanareikin, and A. A. Kostylev,
[28] Order of Magnitude Ahead. Accessed: Jan. 15, 2019. [Online]. Avail- “Classification and Recognition of Targets,” in Radar Target Character-
able: https://wipl-d.com/support/the-order-of-magnitude-ahead/ istics: Measurements and Applications, L. Y. Astanin, A. A. Kostylev,
[29] G. Gentili, V. Tesi, M. Linari, and M. Marsili, “A versatile microwave Y. S. Zinoviev, and A. Y. Pasmurov, Eds. Boca Raton, FL, USA: CRC
plethysmograph for the monitoring of physiological parameters,” IEEE Press, 1994, ch. 9, pp. 337–368.
Trans. Biomed. Eng., vol. 49, no. 10, pp. 1204–1210, Oct. 2002. [56] S. Villeval, I. Bilik, and S. Z. Gurbuz, “Application of a 24 GHz FMCW
[30] C. Gabriel, “Compilation of the dielectric properties of body tissues automotive radar for urban target classification,” in Proc. IEEE Radar
at RF and microwave frequencies.,” Dept. Phys., King’s Coll London, Conf., May 2014, pp. 1237–1240.
London, U.K., Brooks Air Force Tech. Rep. AL/OE-TR-1996-0037, [57] G. J. Mendis, J. Wei, and A. Madanayake, “Deep learning cognitive radar
1996. for micro UAS detection and classification,” in Proc. Cognit. Commun.
[31] M. Chen, C.-C. Chen, S. Y.-P. Chien, and R. Sherony, “Artificial skin for Aerosp. Appl. Workshop (CCAA), Jun. 2017, pp. 1–5.
76–77 GHz radar mannequins,” IEEE Trans. Antennas Propag., vol. 62, [58] F. Chollet, Deep Learning With Python, 1st ed. Greenwich, CT, USA:
no. 11, pp. 5671–5679, Nov. 2014. Manning Publications, 2017.
[32] A. Bhattacharya, B. Pourseyed, and R. G. Vaughan, “Comparison of [59] K. J. Piczak, “Environmental sound classification with convolutional
antenna patterns from active and passive measurements,” in Proc. IEEE neural networks,” in Proc. IEEE 25th Int. Workshop Mach. Learn. Signal
Int. Symp. Antennas Propag. USNC/URSI Nat. Radio Sci. Meeting, Process. (MLSP), Sep. 2015, pp. 1–6.
Jul. 2018, pp. 365–366. [60] V. Boddapati, A. Petef, J. Rasmusson, and L. Lundberg, “Classifying
[33] W. W. Weinstock, “Radar Cross-Section Target Models,” in Mod- environmental sounds using image recognition networks,” Procedia
ern Radar: Analysis, Evaluation, and System Design, 1st ed., Comput. Sci., vol. 112, pp. 2048–2056, Jan. 2017.
R. S. Berkowitz, Ed. Hoboken, NJ, USA: Wiley, Dec. 1965, [61] H. Sinha and P. K. Ajmera, “Interweaving convolutions: An application
pp. 546–574. to audio classification,” BITS Pilani, Pilani, India, Tech. Rep., 2018.
[34] R. Vaughan and J. B. Andersen, Channels, Propagation and Antennas [Online]. Available: https://www.kdd.org/kdd2018/files/deep-learning-
for Mobile Communications, vol. 50. London, U.K.: Institution of day/DLDay18_paper_30.pdf
Engineering and Technology, 2003. [62] C. Szegedy et al., “Going deeper with convolutions,” in Proc. IEEE
[35] B. R. Mahafza, Radar Systems Analysis and Design Using MATLAB, Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2015, pp. 1–9.
2nd ed. Boca Raton, FL, USA: CRC Press, 2005, Ch. 1, 2, 3, 9 and 13. [63] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and
[36] D. K. Barton, Radar Equ. for Modern Radar, 1st ed. Norwood, MA, R. R. Salakhutdinov, “Improving neural networks by preventing co-
USA: Artech House, Dec. 2012, Ch. 2, 6, and 10. adaptation of feature detectors,” 2012, arXiv:1207.0580. [Online]. Avail-
[37] G. L. Charvat, Small and Short-Range Radar Systems, 1st ed. able: https://arxiv.org/abs/1207.0580
Boca Raton, FL, USA: CRC Press, Dec. 2014, Ch. 1, 2, 3, 8 and 9. [64] The Re-Treat Doggy Daycare. Accessed: Jun. 15, 2019. [Online].
[38] I. V. Komarov and S. M. Smolskiy, Fundamentals of Short-Range FM Available: https://theretreatdoggydaycare.com/
Radar. Norwood, MA, USA: Artech House, 2003. [65] Good Karma Doggy Daycare Burnaby. HOME. Accessed: Jun. 15, 2019.
[39] G. L. Charvat, J. H. Williams, A. J. Fenn, and J. S. K. Steve, [Online]. Available: https://goodkarmadoggiedaycare.com/
“Projects,” MIT Lincoln Lab., Massachusetts Inst. Technol., Cambridge, [66] Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality
MA, USA. [Online]. Available: https://ocw.mit.edu/resources/res-ll- assessment: From error visibility to structural similarity,” IEEE Trans.
003-build-a-small-radar-system-cap able-of-sensing-range-doppler-and- Image Process., vol. 13, no. 4, pp. 600–612, Apr. 2004.
synthetic-aperture-radar-imaging-january-iap-2011/projects
[40] XeThru. XeThru Radar in Action. Accessed: Jul. 13, 2018. [Online].
Available: https://www.xethru.com/
[41] Ancortek. About Ancortek. Accessed: Jul. 13, 2018. [Online]. Available:
https://www.ancortek.com/
[42] Imec R&D.. Imec R&D, Nano Electronics and Digital Technologies.
Accessed: Jul. 13, 2018. [Online]. Available: https://www.imec-int.com/
[43] K. Lertsakwimarn, C. Phongcharoenpanich, and T. Fukusako, “A ring
antenna for dual-sense circular polarization,” in Proc. IEEE Int. Conf.
Comput. Electromagn. (ICCEM), Feb. 2016, pp. 194–196.
[44] P. Poshala, K. Rushil, and R. Gupta, “Signal chain noise figure analysis,”
Texas Instrum., Dallas, TX, USA, Appl. Rep. SLAA652, Oct. 2014.
[45] A. G. Stove, “Linear FMCW radar techniques,” IEE Proc. F-Radar
Signal Process., vol. 139, no. 5, pp. 343–350, Oct. 1992.
[46] O. Toker and M. Brinkmann, “A novel nonlinearity correction algorithm Abhijit Bhattacharya (Student Member, IEEE)
for FMCW radar systems for optimal range accuracy and improved received the bachelor’s and master’s degrees in
multitarget detection capability,” Electronics, vol. 8, no. 11, p. 1290, electronics and communications engineering in
Nov. 2019. India and the Ph.D. degree from Simon Fraser
[47] K. Siddiq, R. J. Watson, S. R. Pennock, P. Avery, R. Poulton, and University (SFU), BC, Canada. He spent sev-
B. Dakin-Norris, “Phase noise analysis in FMCW radar systems,” in eral years in the wireless communications indus-
Proc. Eur.Radar Conf. (EuRAD), Sep. 2015, pp. 1523–1526. try prior to his doctoral endeavors, designing
[48] D. E. Barrick, “FM/CW radar signals and digital processing,” Nat. antennas and radio frequency (RF) components
Ocean. Atmos. Admin. Boulder Co Wave Propag. Lab., Silver Spring, for applications ranging from cellular base sta-
MD, USA, Tech. Rep. ERL 283-WPL 26, 1973. tions and wireless backhauls to portable mobile
[49] A. Goldsmith, Wireless Communications, 1st ed. Cambridge, U.K.: devices. He is currently a Staff Engineer (RF and
Cambridge Univ. Press, 2005, ch. 2, pp. 25–40. Antenna design) with Sierra Wireless Inc., BC. His current research inter-
[50] J. R. Wait, Electromagnetic Wave Theory, 1st ed. New York, NY, USA: ests include radars, deep-learning for radar and electromagnetic applica-
Harper & Row, 1985, ch. 4, pp. 148–158. tions, antenna theory, and design and testing for MIMO communications.

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.
 BHATTACHARYA AND VAUGHAN: DEEP LEARNING RADAR DESIGN FOR BREATHING AND FALL DETECTION 5085

Rodney Vaughan (Life Fellow, IEEE) (PEng) is technology, industrial electronics, signal theory and processing, sampling
the Sierra Wireless Professor in Communications theory, diversity design, antennas for personal, cellular, and satellite com-
with Simon Fraser University (SFU), BC, Canada. munications, and “large-N” MIMO system design. His current research is
His university training was in New Zealand and in propagation, signal processing with deep learning, acoustics, commu-
Denmark. Most of his career has been with indus- nications techniques, and the theory, design, and evaluation of antennas.
trial design, and he joined academia in 2003. Recent antenna projects include on-chip antennas, bicycle frame anten-
Before 2003, he worked with the New Zealand nas, arrays for satellite SAR, and industrial food heating. He has edited
Post Office (later called Telecom New Zealand for the IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION (compact
Limited and Kora Aotearoa), the NZ Department elements, arrays, mobile antennas, multipath, and diversity/MIMO tech-
of Scientific and Industrial Research, and Indus- niques), and has guest-edited a variety of Special Issues. He is a Fellow
trial Research Limited (IRL). He covered a wide of the BC Advanced System Institute in 2004, an URSI Correspondent,
range of mechanical and electrical projects, including heating and ventila- and continues as the New Zealand URSI Commission B representative.
tion, communications network analysis, and developed microprocessor In 2006, he served on the international panel for reviewing the funding
and DSP technology for equipment ranging from abattoir hardware to and knowledge base for ICT in the U.K., and in 2011 for the Canadian
real-time acoustic noise cancellation. He was an URSI Young Scientist Communications Research Centre and also for the ICT Centre of the
in 1982 for fields and waves and in 1983 for electromagnetic theory. CSIRO. He was a multiterm IEEE Distinguished Lecturer for the VT Soci-
He developed research, programmes and personnel in communications ety, and he has been a Co-General Chair for major IEEE conferences.

Authorized licensed use limited to: UNIVERSITAT POLITECNICA DE CATALUNYA. Downloaded on July 24,2024 at 10:40:18 UTC from IEEE Xplore. Restrictions apply.