Vital Signs Monitoring of Multiple People using a

FMCW Millimeter-Wave Sensor

Adeel Ahmad, June Chul Roh, Dan Wang, Aish Dubey

Texas Instruments Incorporated
12500 TI Boulevard,
Dallas, Texas, USA

Abstract—In this paper, we describe non-contact monitoring II. BACKGROUND

of vital signs of multiple people using a frequency modulated
continuous wave (FMCW) sensor. We use the inherent range- A. Signal Model
gating ability of the FMCW waveforms and multiple receive The FMCW signal model is described in detail in several
channels to separate objects in range-azimuth plane. We sub-
previous studies [4]. Briefly, a periodic linearly-increasing
sequently utilize the fact that body surface movements due to
physiological motions modulates the phase of the received radar frequency ramps are transmitted given by
signal and can be further processed to extract the breathing and
heart-rate. Range-gating and beamforming techniques allow the B 2
signal of interest to be isolated from surrounding clutter, however xT (t) = AT cos(2πfc t + π t + φ(t)) (1)
Tc
several challenges such as random body movements need to be
addressed before radar-based non-contact measurements can be where AT is the transmitted power and φ(t) is the phase
deployed in real-world settings.
noise from the transmitter. Some relevant chirp parameters are
I. I NTRODUCTION defined as below

Currently the main focus of mm-wave radar applications fc : Chirp starting frequency
are geared towards the automotive market [1], however sev- B: Bandwidth of the chirp
eral potential application areas within the broad industrial
and health-care domain are also under active investigation. Tc : Chirp duration
Millimeter-wave radars are currently being explored for use Tf : Fast-time-axis ADC sampling interval
in numerous medical and diagnostic applications ranging from
cancer imaging to glucose monitoring [2]. One emerging Ts : Slow-time-axis sampling interval
application area is remote non-contact monitoring of human The radar receives a scaled and shifted (by α and td )
vital signs [3]. This can be important in several scenarios such version of the transmitted signal given by
as monitoring a room full of sleeping elderly or inebriated
people to diagnose life threatening situations at low cost; in
critically-ill patients where non-contact measurements might B
be more desirable such as in neonatal units for babies and for xR (t) = αAT {cos(2πfc (t − td ) + π (t − td )2 + φ(t − td )
Tc
burn victims; for driver monitoring to detect driver drowsiness (2)
and sleepiness to avoid auto accidents; in detecting humans
where td = 2R(t)/c is the range-dependent time delay
who are either buried in emergency accident situations or
from a given object at radial range R(t) . The received signal
hiding to resolve a security situation etc.
is mixed with a replica of the transmitted signal and after I/Q
In this paper, we demonstrate the feasibility of measur- mixing the signal can be approximated as
ing human vital-signs of multiple people using a frequency B
td ]t+2πfc td +π B
t2d +Δφ(t))
modulated continuous wave (FMCW) CMOS mm-wave sensor y(t) = AR ej(2π[ Tc Tc

operating from 76-81 GHz. We utilize the frequency modulated = AR ej(2πfb t+φb (t)+Δφ(t)) (3)
waveform and measurements over multiple receive channels to
separate objects in the range-azimuth angle plane and further where AR is the received signal power, fb is the beat
process the phase signal to extract the breathing-rate and heart- frequency given by
rate. We also demonstrate that these applications could be
reliably supported by an integrated solution which helps in 2BR(t)
fb = , (4)
optimizing the size, power consumption and cost of the safety cTc
critical monitoring solutions. In section 2 of this paper, we and
briefly describe the vital signs signal model from a FMCW B 2
φb (t) = 2πfc td + π td (5)
radar perspective and measurement of the displacement sen- Tc
sitivity of our system. In section 3, we explain the main
is the phase.
processing steps to extract the breathing and heart-rate from
multiple people within the radar field-of-view followed by a The residual phase noise is (Δφ(t) = φ(t) − φ(t − 2R/c))
discussion on the current challenges and future work. and can be negelcted for short-range radar applications due to

978-1-5386-4167-5/18/$31.00 ©2018 IEEE 1450

the range-correlation effect. Additionally, the term π TBc t2d can system. We define phase sensitivity as the RMS value of the
also be neglected in φb as it is negligibly small in practical phase fluctuations of a static object. Phase sensitivity of the
scenarios. system would determine the smallest observable/measurable
change in phase that the system can measure and is related to
The beat signal after I/Q sampling can be expressed for the the displacement sensitivity by σR = 4πλ
σφ
nth ADC sample and mth chirp as
4π
y[n, m] = AR ej(2πfb nTf + λ R(nTf +mTs )) (6)
In a system with multiple receivers, an additional phase shift
will occur due to the relative distances between the various
receiver elements. For an object at angle, the signal received
at receiver i that is placed at a relative distance di (assuming
a uniform linear array) is
4π 2π
y[n, m, i] = AR ej(2πfb nTf + λ R(nTf +mTs )+ λ di sin θ) (7)
If multiple targets are present within the radar field-of-view,
then the received signal is additive and for L number of targets
at (Rl , θl ) is given as below
Fig. 1. Noise model for determining the phase sensitivity of the system

L
j(2πfbl nTf + 4π 2π
λ Rl (nTf +mTs )+ λ di sin θl )
yT [n, m, i] = ARl e
l=1 To derive, the theoretical curve for phase/displacement
(8) sensitivity we use a simplified noise model shown in Figure
th
Based on the above equation, the phase shift at the i receiver 1. The measured signal can be modelled as the summation
is of the ideal signal S and noise N  . Based on Figure 1,
4πRl (nTf + mTs ) 2π the phase variations φ due to noise in the measured signal
φb,i (nTf + mTs ) = + di sin θl (9)
λ λ can be approximated as tan φ ≈ φ = |N ||S| sin ϕr
. Under
Body surface displacement due to vital signs is of small the assumption that noise has a random phase ϕr uniformly
amplitudes (< 10 mm) and low frequency (< 4 Hz). This distributed between 0 and 2π, the phase variance is given by
2
implies that there would be no appreciable change in phase σφ2 = 12 |N |
|S 2 | =
1
2 and the RMS value is σφ = √
2(SN R)
1
2(SN R)
during the chirp time (fast-time axis) and measuring phase
changes induced between successive chirps (slow-time axis)
would be sufficient. If the object is at a nominal range Rl,o
then
Rl (nTf + mTs ) ≈ Rl,o + Rl (mTs ) (10)
4πRl,o 4πRl (mTs ) 2π
φb,i (mTs ) = + + di sin θl (11)
λ λ λ
4πRl (mTs )
= φc,i (Rl,o , θ) + (12)
λ
where φc,i (Rl,o , θl ) would be a constant for an object
that remains in the same range-angle bin i.e. (Rl,o , θl ). In
vital signs measurement, we are mainly interested in the
relative displacement, Rl (mTs ) and hence the constant phase Fig. 2. Phase sensitivity measurements. Phase variations were measured at
term is not considered, however, it can be seen that any the range-bin corresponding to a static target.
changes in (Rl,o , θl ) during the course of the measurements
can significantly affect the results. The phase changes across
the receivers are however utilized to monitor and analyze the We experimentally determined the phase sensitivity by
vital signs for people that are in separately identifiable range- measuring the phase variation across the object range-bin as a
angle bins. function of time for a static corner reflector placed at a fixed
distance. The SNR at the object was varied by changing the
transmit power. As can be seen from the plot in Figure 2 that
B. Displacement Sensitivity
the experimental values follow the theoretical values and at
Its apparent from the expression of φb that a smaller λ SNR > 40 dB, the phase sensitivity is < 7 milli-radians which
will generate a larger phase change for a given displacement corresponds to a displacement sensitivity of ≈ 2 microns. This
R(t). This explains why 77 GHz, millimeter-wave radars by gives us confidence that the system has the ability to measure
virtue of their smaller wavelengths (≈3.9 mm), have greater tens of microns of displacements for heart-beat measurements.
sensitivity in measuring small displacements that can be as Further, the transmit power used to create such an antenna
low as tens of microns due to heart beats. As we are using system can be as low as 2mW for the CMOS setup and is
the phase to measure the vibrational amplitude and frequency, less than the safe exposure limits for human body towards
hence it is useful to characterize the phase sensitivity of the non-ionizing radiation [5].

978-1-5386-4167-5/18/$31.00 ©2018 IEEE 1451

 TABLE I. PARAMETERS USED FOR VITAL SIGNS MEASUREMENTS
by T . However, as the vital signs signals are non-stationary,
System Parameters System Parameters the observation time is usually limited to 8 – 15 seconds due
Starting Frequency 77 GHz Max. Range 4.3 m to the inherent time-frequency tradeoffs.
Slow-axis Sampling 20 chirps/sec Range Resolution 4.3 cm
Chirp Duration 50 sec Azimuth Angular resolution 15 B. Extraction of Vital Signs Signature from Multiple People
ADC Sampling Rate 2 Msps Transmit Power 10 dBm
As a first step, we need to identify the range-angle bins
corresponding to the person within the radar field-of-view.
III. P ROCESSING STEPS Once the appropriate range-angle bins have been identified
then the processing steps of Figure 4 can be applied to
In this section we describe the mm-wave sensor parameters extract the vital signs waveform and Figure 6 to estimate the
and signal processing steps involved in the extraction of the breathing-rate and heart-rate. The range-angle bin tracker runs
breathing-rate and heart-rate. at a user-defined time interval (e.g. every 30 seconds) and its
purpose is to identify the angles of the objects of interest
A. Radar Signal Acquisition Parameters and then compute appropriate beam-forming weights based
on these angles. Standard radar signal processing chain as
The mmwave sensor used in these experiments was a single
outlined below is used to extract the vital signs of multiple
TI-mmWave IWR1443 sensor operating at 77-81 GHz with 4
people within the field-of-view
GHz chirp bandwidth [6]. The system has 3 transmit (TX)
antennas and 4 receive (RX) antennas as shown in Figure 1) : A FMCW is transmitted over multiple TX antennas in
3. In order to increase the azimuth angular resolution, the a time division multiplexing (TDM) manner and received over
chirps were transmitted in a time division multiplexing MIMO several receive (RX) channels effectively forming multiple TX-
configuration by transmitting sequentially through Tx-1 and RX pairs. We use a MIMO configuration to achieve better
Tx-3. This results in an 8 element virtual array that gives a azimuth angular resolution
theoretical azimuth resolution of 15o . Some key system and
chirp parameters are shown in Table 1 and described below. 2) : A range-FFT is performed for each of these TX-RX
pairs. This would allow objects that occupy different range
bins to be separated out from one another.
3) : A second FFT is taken along the RX dimension (or
other direction of arrival estimation algorithms can be used)
for each of the range bins forming a range-azimuth plane. This
Fig. 3. Antenna pattern on the IWR1443 EVM. Chirps were transmitted
would separate objects that fall in the same range bin but are
sequentially through Tx-1 and Tx-3 using time division multiplexing resulting at different angles from one another.
in an 8 element virtual array
4) : Objects are detected in the range-azimuth plane using
1) Slow-time axis sampling rate: The slow-axis sampling an object detection scheme, such as CFAR (constant false
rate (chirps/second) should satisfy (a) Nyquist criteria i.e. the alarm rate).
sampling rate should be twice the maximum frequency in the 5) : For each detected object in the range-azimuth plane,
signal and to prevent any aliasing of the noise. Given that we measure the phase of the particular range-azimuth angle bin
our frequency of interest is within 0.1 – 4 Hz we chose a over the slow-time axis and only retain those range-angle bins
sampling rate of 20 Hz to ensure sufficient oversampling of whose phase variation over time exceeds a certain threshold
the signal. Moreover, (b) the sampling rate should be high (to filter out the reflections from static objects).
enough to prevent excessive phase wrapping. For a sinusoidally
vibrating object A sin(2πfm t) this can be satisfied by choosing 6) : Once the appropriate range-angle bins have been iden-
a slow-axis sampling rate such that Fs > 8πfλm A where A is tified, beamforming weights are computed for each identified
the amplitude and fm is the vibration frequency. angular position k and applied to the ADC data

2) Chirp Duration: The chirp duration chosen for these 

IRx
yk [n, m] = y[n, m, i]wki (13)
experiments was 50 us. In general higher the chirp duration,
i=1
the better the SNR and displacement sensitivity.
3) Bandwidth: Its well known that the bandwidth of the 7) : After the Range-FFT, the phase values are extracted
chirp is directly related to the range resolution of the system. from the identified range-bins and subsequently passed through
Although the bandwidth does not directly impact the vital a signal processing chain (shown in Figure 6) to estimate the
signs estimation accuracy however it can significantly influence vital signs.
the overall performance as a higher range resolution system
enables the separation of the area of interest (e.g. chest) from C. Body Displacements Induced by Physiological Motion
other limb movements and also allows better separation of one
persons vital signs from another. At GHz frequencies, displacement of the body surface is
the dominant physical mehcanism that enables non-contact vi-
4) Observation Time: The larger the observation time T , tal signs measurements [7]. The body surface displacement due
the better the frequency and amplitude estimation. According to breathing and heart-beat will modulate both the magnitude
to the Cramer Rao Lower Bounds (CRLB), the frequency esti- AR and phase φb of the received signal. The fluctuations in AR
mation improves on the order of T 3 and amplitude estimation are most likely due to the variations in the radar cross-sectional

978-1-5386-4167-5/18/$31.00 ©2018 IEEE 1452

Fig. 4. Signal processing chain for separting multiple targets. Range-angle bin
tracker finds the angular positions of the objects. Beamforming is subsequently
performed using the weights corresponding to these angles followed by the
range-FFT

(RCS) area of the human body due to displacements caused

by breathing and heart-beat. In this work, we are utilizing the Fig. 5. Chest displacement and its frequency content (a) Unwrapped phase
phase φb to extract the vibration pattern of the human chest as and its frequency content (b) Unwrapped phase after the phase difference
operation. Note that the heart-rate frequencies become more obvious. (c)
the phase changes are directly related to the relative changes Another dataset of chest displacement (after phase unwrapping and difference)
in displacement. We also assume that the subject stays within showing amplitude modulation of the heart-beat signal. fbr is the breathing-
the same angle-range bin during the measurements. rate while fh is the heart-rate.

In general, the chest displacement waveforms can have

significant variations depending on the spatial location of mea- D. Signal Processing for Vital Signs Estimation
surements, inter-person variations, state of health etc. making
it difficult to come up with a general model characterizing Vital signs are estimated using a running window of T ≈
the vital signs waveform. Some typical examples of the phase 15 seconds. The algorithm for vital-sign estimation consists of
changes induced by chest displacements are shown in Figure the following main modules
5. These were measured from a subject sitting stationary on 1) Phase extraction and unwrapping: The phase values are
a chair to minimize random body movement. From Figure computed using an arc tangent operation and unwrapped to
5(a), the larger amplitude displacements are due to breathing obtain the actual displacement profile.
while the small oscillations on top of the breathing signal
Q
are due to heart-beats. Its immediately apparent that the chest φ(m) = unwrap[tan−1 ( )] (14)
displacement is dominated by the breathing signal which can I
be an order of magnitude (2x – 40x) higher than the heart-beat Subsequently, the phase differences between successive un-
signal. From the frequency spectrum (right column) it can be wrapped phases are computed which helps in removing any
seen that both the breathing ( fbr ) and the heart-beat (fh ) can phase drifts and in suppressing the breathing and its harmonics.
have harmonic frequencies as these signals are rarely purely
sinusoidal. 2) Noise removal: The un-wrapped differential phase a(m)
might be corrupted by several noise-induced phase wrapping
In order to minimize the impact of the larger ampli- errors especially if the phase values are close to -π or π. This
tude breathing signal on the heart-rate estimate we take the impulse-like noise is removed by computing a forward a(m)−
phase difference between successive phase samples followed a(m + 1) and backward a(m) − a(m − 1) phase difference for
by band-pass filtering. Although this does not completely each a(m) and if these exceed a certain threshold then a(m)
eliminate the impact of breathing and its harmonics on heart- is replaced by an interpolated value.
rate estimate but nevertheless helps in enhancing the heart-
3) Separation of breathing and heart-beat: The breathing
rate signal. An example is shown in Figure 5(b) where the
and heart-rate occupy different frequency bands and hence
heart-rate frequency and its harmonics are more apparent. At
can be separated by frequency filtering. A 4th order IIR-
times, the breathing harmonic frequencies can overlap with the
cascaded Bi-quad filter was used to bandpass the signal into
cardiac-region spectrum (0.8 – 4.0 Hz) significantly affecting
the cardiac-region spectrum [0.8 - 4.0] Hz and the breathing-
the heart-rate measurements. We also note that the heart-
region spectrum [0.1 - 0.5] Hz.
beat signal can be modulated by random body movement and
breathing pattern. In the dataset shown in Figure 5(c), the 4) Motion corrupted segment removal: This block reduces
amplitude modulation of the heart-beat signal can be observed the impact of any large amplitude movements on the heart-rate
which results in the presence of side-bands i.e. (fh ± fbr ). The estimates. The energy in the cardiac waveform is computed
presence of all these frequency components makes picking the for a window size of 1 sec and if the energy E within this
peak corresponding to the heart-rate difficult, but not impossi- segment exceeds a user-defined threshold (E > ET h ), then
ble especially if intelligent DSP algorithms are employed that all the samples in that segment are discarded from the time-
integrate and sift the data over many seconds of operations. domain cardiac waveform.

978-1-5386-4167-5/18/$31.00 ©2018 IEEE 1453

 5) Windowing and gain control: Prior to spectral estima- IV. R ESULTS
tion, windowing of the signal and gain control is applied. The
purpose of the gain control step is to reduce the impact of any In the results shown in Figure 7, we placed two targets
large amplitude movements on the vital signs estimates. at the same range-bin but at different angles. Target 1 is a
corner reflector vibrating with a fixed frequency (0.85 Hz)
6) Spectral estimation: The separated breathing and heart and amplitude (50 um) while target 2 is a person periodically
rate waveforms are passed through a spectral estimation mod- breathing normally and holding the breath. As expected, when
ule. The spectral estimation is done through the FFT algorithm. beamforming is not applied the resultant signal will be a vector
summation of the two signals and the individual displacements
7) Inter-peak distance: Heart-rate and breathing-rate are cannot be discerned as shown in Figure 7 (a). However, after
also estimated by finding the inter-peak distances in their applying beam-forming the displacements from target 1 and
respective time-domain waveforms. Two thresholds i.e. min- target 2 can be clearly seen in Figure 7 (b). The separation
imum peak distance (Pmin ) and maximum peak distance needed between targets would depend on the range and angle
(Pmax ) are defined based on the sampling rate and allowed resolution of the system which in our case is 4 cm and 15o
frequency range (for heart-rate this would be 0.8 - 2.0 Hz). The respectively.
first peak in the waveform is chosen as a valid peak and the
next valid peak is chosen such that the distance between the
current peak and the previous valid peak is within the interval
[Pmin Pmax ]. Additionally peak can be rejected based on if
their amplitudes are too high or too low based on the expected
vital-signs signal amplitudes. After isolating the valid signal
peaks, an estimate is made based on the average of the inter-
peak distances for all the valid peaks. The inter-peak distance
is used as a fallback estimate of the vital signs if the confidence
metric on the FFT-based estimate is too low.

8) Breathing-Rate estimate: The breathing rate is chosen

based on the frequency of the maximum-peak within the
breathing-region spectrum. A confidence metric is computed
as the ratio of the signal power of the maximum-peak (and Fig. 7. Two targets placed at the same range bins but at different angles.
Target 1 is a corner reflector vibrating with f = 0.85 Hz with an amplitude of
some frequency bins around it) over the remaining frequency 50 um while target 2 is person periodically holding his breath and breathing
bins in the breathing-region spectrum. If confidence metric is normally. (a) Range-profile from a single antenna. The displacement profiles
below a certain threshold, then the inter-peak distance based of the two targets are non-discernible (b) Range-azimuth plane showing the
estimate is chosen as the breathing-rate. two targets and the corresponding displacement waveforms after beamforming
is used to separate the targets.
9) Heart-Rate estimate: The maximum peak in the cardiac
spectrum may not necessarily correspond to the heart-rate. In Figure 8, we show vital signs measurements taken
Breathing harmonics, noise sources, body movements, can simultaneously from two people sitting at the same radial
result in larger magnitude peaks than the actual heart rate peak. distance (≈ 1 m) from the radar with an angular separation
Therefore we estimate the heart-rate using a density-based of ≈ 60o . They were asked to alternate between normal
approach that is: find all peaks in the cardiac spectrum and breathing and holding their breath. The separation of the chest
retain top N peaks; remove the peaks (from the top N peaks) displacement from each of these people can be seen in the
that correspond to breathing harmonic; place the remaining plots (while one person was holding the breath the other was
peaks in the circular buffer; accumulate the peaks for T breathing normally).
seconds; use a clustering algorithm such as a dBscan algorithm
to divide the accumulated peaks into clusters; determine the
cluster with the maximum number of peaks; and choose the
median peak value of the determined cluster as valid heart-rate
value.

Fig. 8. Simultaneous measurement of vital signs of two people sitting at

the same radial distance from the radar but at different azimuth angles. The
subjects were asked to alternate between normal breathing and holding their
breath and while one person was holding the breath the other was breathing
normally.
Fig. 6. Block diagram of vital signs estimation algorithm
Figure 9 shows an example of the breathing- and heart-

978-1-5386-4167-5/18/$31.00 ©2018 IEEE 1454

 [7] O. Aardal, Y. Paichard, S. Brovoll, T. Berger, T. S. Lande, and S. E.
Hamran, “Physical working principles of medical radar,” Biomedical
Engineering, IEEE Transactions on, vol. 60, no. 4, pp. 1142–1149,
2013.
[8] F. Adib, H. Mao, Z. Kabelac, D. Katabi, and R. Miller, “Smart homes
that monitor breathing and heart rate,” in Proc. 33rd Annual ACM
Conference on Human Factors in Computing Systems, Seoul, Republic
of Korea, 2015.
[9] B. Schleicher, I. Nasr, A. Trasser, and H. Schumacher, “Ir-uwb radar
demonstrator for ultra-fine movement detection and vital-sign monitor-
ing,” Microwave Theory and Techniques, IEEE Transactions on, vol. 61,
Fig. 9. Example signal waveform (a) chest displacement. Bandpass filters no. 5, pp. 2076–2085, 2013.
applied to the chest displacement to obtain the (b) Breathing waveform and (c)
heart-beat waveform. (d) Breathing-rates and (e) Heart-rates estimated from [10] G. Shafiq and K. C. Veluvolu, “Surface chest motion decomposition
the corresponding waveforms. for cardiovascular monitoring,” Sci. Rep., vol. 4, 2014. [Online].
Available: http://dx.doi.org/10.1038/srep05093
[11] I. Mostafanezhad, E. Yavari, O. Boric-Lubecke, V. M. Lubecke, and
D. P. Mandic, “Cancellation of unwanted doppler radar sensor motion
waveforms and breathing-rate and heart-rate obtained after using empirical mode decomposition,” Sensors Journal, IEEE, vol. 13,
applying the signal processing chain in Figure 6. no. 5, pp. 1897–1904, 2013.
[12] L. Changzhi and L. Jenshan, “Random body movement cancellation in
doppler radar vital sign detection,” Microwave Theory and Techniques,
V. C ONCLUSION AND FUTURE WORK IEEE Transactions on, vol. 56, no. 12, pp. 3143–3152, 2008.
[13] I. V. Mikhelson, P. Lee, S. Bakhtiari, T. W. Elmer, A. K. Katsaggelos,
Radar based non-contact vital sign monitoring has been and A. V. Sahakian, “Noncontact millimeter-wave real-time detection
demonstrated using several radar technologies such as UWB, and tracking of heart rate on an ambulatory subject,” Information
CW Doppler radar and FMCW [8], [9]. While all these meth- Technology in Biomedicine, IEEE Transactions on, vol. 16, no. 5, pp.
ods work fine under controlled environment and conditions, 927–934, 2012.
several challenges need to be addressed to make these sys- [14] G. Changzhan, W. Guochao, L. Yiran, T. Inoue, and L. Changzhi,
tems reliable and robust enough to be deployed in real-world “A hybrid radar-camera sensing system with phase compensation for
random body movement cancellation in doppler vital sign detection,”
settings. These challenges include dealing with the impact of Microwave Theory and Techniques, IEEE Transactions on, vol. 61,
random body and sensor movement, effective cancellation of no. 12, pp. 4678–4688, 2013.
breathing harmonics and robust measurements in a clutter rich [15] J. Vicente, P. Laguna, A. Bartra, and R. Bailon, “Drowsiness detection
environment etc. In general, solving for these challenges will using heart rate variability,” Med Biol Eng Comput, vol. 54, no. 6, pp.
require more signal processing capabilities [10], [11], more 927–37, 2016.
radars [12] or combining radars with other sensing modalities [16] S. Suzuki, T. Matsui, H. Imuta, M. Uenoyama, H. Yura, M. Ishihara,
[13], [14]. For more robust measurements and motion compen- and M. Kawakami, “A novel autonomic activation measurement
method for stress monitoring: non-contact measurement of heart
sation it might be beneficial to take measurements of the same rate variability using a compact microwave radar,” Med Biol Eng
set of people from multiple overlapping view-points. This can Comput, vol. 46, no. 7, pp. 709–714, 2008. [Online]. Available:
be done for example by connecting multiple CMOS mm-wave http://dx.doi.org/10.1007/s11517-007-0298-3
sensors to a TDA3x embedded processor device. In addition to [17] M. Zhao, F. Adib, and D. Katabi, “Emotion recognition using
vital signs, mmwave sensors can potentially be used to measure wireless signals,” in Proceedings of the 22Nd Annual International
Conference on Mobile Computing and Networking, ser. MobiCom ’16.
drowsiness [15], stress levels [16] and human emotions [17]. New York, NY, USA: ACM, 2016, pp. 95–108. [Online]. Available:
Continued experimentation and research into these systems http://doi.acm.org/10.1145/2973750.2973762
is still necessary for critical safety and emergency response
applications that have been listed out in section 1 of the paper.

R EFERENCES
[1] J. Hasch, E. Topak, R. Schnabel, T. Zwick, R. Weigel, and C. Wald-
schmidt, “Millimeter-wave technology for automotive radar sensors in
the 77 ghz frequency band,” IEEE Transactions on Microwave Theory
and Techniques, vol. 60, no. 3, pp. 845–860, 2012.
[2] F. Topfer and J. Oberhammer, “Millimeter-wave tissue diagnosis: The
most promising fields for medical applications,” Microwave Magazine,
IEEE, vol. 16, no. 4, pp. 97–113, 2015.
[3] C. Gu, “Short-range noncontact sensors for healthcare and
other emerging applications: A review,” Sensors (Basel,
Switzerland), vol. 16, no. 8, p. 1169, 2016. [Online]. Available:
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC5017335/
[4] G. M. Brooker, “Understanding millimetre wave fmcw radars,” in Proc.
1st International Conference on Sensing Technology, Palmerston North,
New Zealand, 2005.
[5] Non-ionizing radiation safety manual. [Online]. Avail-
able: https://ehs.berkeley.edu/laser-safety/non-ionizing-radiation-safety-
manual
[6] Texas instruments mmwave sensors. [Online]. Available:
https://www.ti.com/sensing-products/mmwave/overview.html

978-1-5386-4167-5/18/$31.00 ©2018 IEEE 1455