Remote vital signs monitoring using

a mm-wave FMCW radar

by

Mostafa Alizadeh

A thesis
presented to the University of Waterloo
in fulfillment of the
thesis requirement for the degree of
Master of Science
in
Electrical and Computer Engineering

Waterloo, Ontario, Canada, 2019

© Mostafa Alizadeh 2019

 Author’s Declaration

I hereby declare that I am the sole author of this thesis. This is a true copy of the thesis,
including any required final revisions, as accepted by my examiners.

I understand that my thesis may be made electronically available to the public.

ii
 Abstract

A vision on the migration from contact standard health monitoring measurement devices
to non-contact measurement technologies has gained a tremendous attention in literature
and in industry. A promising method for realizing the remote measurement of vital signs
is using electromagnetic radars such as frequency modulated continuous wave (FMCW)
radars. However, using these radars has challenges to precisely acquire the respiration
and heart rates. A solution for higher accurate measurement of the vital signs can be the
use of mm-wave frequencies, which gives a high-resolution sensing of displacements in an
environment in the order of sub-mm changes. On the other hand, being in mm-wave bands
increases both hardware and signal processing designs and implementations.
In this work, a mm-wave radar is used to monitor the breathing and the heart rates
as well as their waveforms for further clinical diagnostics. To that end, we established a
complete analysis of the FMCW radars principles by considering hardware impairments.
The analysis considers the effect of antenna coupling, RF cross-talk, stationary clutters,
phase noise, IQ imbalances, and the thermal noise. Also, the effect of the individual
hardware imperfections on the phase quality is shown by simulations and experiments.
The simulations are carried out with a Matlab Simulink model. For the experiments,
Texas Instruments (TI) mm-wave FMCW radars have been used.
To earn insight into vital signs monitoring, different experiments are designed. In the
experiments, the effect of the thermal instability of the RF parts on the phase is shown.
In addition, to mimic the behaviour of the chest vibration due to respiration and the
heartbeats, a two-pendulum system is designed and tested. Particularly, the pendulum
system performance in terms of vibration frequency estimations of the two pendulums
versus distance is then measured. In the simulations, the system performance is obtained
for different signal to noise ratios (SNR) and different phase noise levels, as well as different
stationary clutters.
Finally, to test the TI sensors for different directions to the subjects, Hexoskin smart
garment is used as a reference sensor, which is a reliable commercial product. Our results
show great system improvement in terms of accuracy of the vital signs detection in com-
parison to other similar research. For different sleep positions, the accuracy of HR and BR
are greater than 94% and 96%, respectively. In addition to detecting the vital rates, we
have shown that their waveforms can also be reconstructed by using an adaptive optimum
filter.

iii
 Acknowledgements

First of all, I should thank God whose blessings are endless during this life or Hereafter.
He gave and will give the meaning, the courage, and the power for this work and the future
works as he said: “The effort of one who faithfully strives hard for the life to come, will be
appreciated.”
I would also like to express my sincere gratitude to my advisers professor Safieddin
(Ali) Safavi-Naeini and professor George Shaker for their continuous support of my
Master study and related research, patience, motivation, and immense knowledge. Their
guidance has helped me in all the time of research and writing of this thesis.
Also, I want to thank all my friends who supported me during the completion of this
degree. I am not mentioning their names because they deserve to be in my heart and my
mind all the time.
A special thank to my parents who are constantly supporting me emotionally and
subjectively. Without them, the creation of this thesis was not possible.

iv
 Dedication

This study is wholeheartedly dedicated to my beloved parents, who have been my source
of inspiration and gave me strength when I thought of giving up, who continually provide
their moral, spiritual, emotional, and financial support.

v
Table of Contents

List of Tables ix

List of Figures x

Abbreviations xiii

List of Notations and Symbols xvi

1 Introduction 1
1.1 A big picture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Literature survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Main contributions of the work . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2 Hardware and equipment 12

2.1 Texas Instruments mm-wave radars . . . . . . . . . . . . . . . . . . . . . . 12
2.1.1 Benefits of using IQ demodulator . . . . . . . . . . . . . . . . . . . 13
2.1.2 General parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Hexoskin vest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

vi
3 System model and radar analysis 19
3.1 FMCW radar fundamental principles . . . . . . . . . . . . . . . . . . . . . 20
3.2 Radar equation and received signal power . . . . . . . . . . . . . . . . . . 22
3.2.1 Complex signal analysis . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 DC compensation process . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4 Phase analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.5 Harmonic analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5.1 Optimum filter design . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.5.2 Harmonic processing complexity . . . . . . . . . . . . . . . . . . . . 39
3.5.3 Real phase to complex phase . . . . . . . . . . . . . . . . . . . . . . 40
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

4 System verification with experiments and simulations 42

4.1 Antenna chamber tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2 Stationary tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3 Pendulum tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.1 One-pendulum tests . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.2 Two-pendulum tests . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.4 Matlab simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5 Application on vital signs detection 69

5.1 Signal processing flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2 Experiments’ setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3 Waveforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.4 Respiration and heart rates . . . . . . . . . . . . . . . . . . . . . . . . . . 81

6 Summary, outlook, and future extensions 84

6.1 Future extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

vii
References 86

APPENDICES 92

A TI radar 93

B Matlab codes and the Simulink model 97

viii
List of Tables

2.1 AWR1443 and AWR1642 specifications . . . . . . . . . . . . . . . . . . . . 17

3.1 FCMW radar range and velocity estimations and their min/max detectable
bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.1 Radar configuration for chamber tests . . . . . . . . . . . . . . . . . . . . . 43

4.2 Min/Max value of different parameters . . . . . . . . . . . . . . . . . . . . 44
4.3 Radar configuration for the Simulink model . . . . . . . . . . . . . . . . . 64
4.4 Radar parameter detection valid ranges for the Simulink model . . . . . . . 64

5.1 HR and BR error rates for different sleep positions: the underlined values
are the worst performances . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.2 Estimation accuracy comparison to similar works . . . . . . . . . . . . . . 83

A.1 CLI command description . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

ix
List of Figures

2.1 IQ demodulation vs real demodulation . . . . . . . . . . . . . . . . . . . . 14

2.2 Real mixing: ∆f1 , ∆f2 , ∆f3 are corresponding to the downconverted fre-
quencies of f1 , f2 , f3 , respectively. . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Falling intermodulations of the LNA output in the based band: the dotted
and solid lines are for negative and positive frequencies, respectively. . . . . 16
2.4 Hexoskin and BioRadio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.1 Signal transmission and reception of FMCW chirps . . . . . . . . . . . . . 19

3.2 Maximum range of a target versus minimum required SNR at the based
band. The gain of Tx/Rx antennas are the same as G. . . . . . . . . . . . 23
3.3 Radial velocity with respect to the radar . . . . . . . . . . . . . . . . . . . 26
3.4 Correction of the received complex signal. . . . . . . . . . . . . . . . . . . 30
3.5 Phase jump cases: phase difference of greater than π (left), phase difference
of less than −π (right). The red lines are the unacceptable paths for phase
change while the green lines are the acceptable paths. . . . . . . . . . . . . 32
3.6 Phase unwrapping test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4.1 Antenna chamber . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.2 Different maps for the small still plate . . . . . . . . . . . . . . . . . . . . 45
4.3 Different maps for the large still plate . . . . . . . . . . . . . . . . . . . . . 46
4.4 Different maps for the large high vibrating plate . . . . . . . . . . . . . . . 47
4.5 Phase variation of the oscillating plate . . . . . . . . . . . . . . . . . . . . 48
4.6 Linear phase trend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

x
4.7 Pendulum system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.8 Phase time trace for the pendulum at 2.5 m. . . . . . . . . . . . . . . . . . 52
4.9 Phase spectrograms of a pendulum with vibration of 2.1 Hz at 30 cm and
2.5 m. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.10 f0 estimation error rate for a single pendulum. . . . . . . . . . . . . . . . . 54
4.11 Signal constellation with a shifted origin . . . . . . . . . . . . . . . . . . . 55
4.12 Phase modulation with a single pendulum before and after DC cancellation 55
4.13 Phase spectrograms before and after DC offset cancellation . . . . . . . . . 56
4.14 PSD of the phase signal before and after DC compensation . . . . . . . . . 56
4.15 Phase spectrograms of the two pendulums at 85 cm and 2.5 m . . . . . . . 58
4.16 Frequency estimator error rate for two-pendulum and single-pendulum ex-
periments at different ranges: f0 , f1 are for p1 and p2, respectively. . . . . 59
4.17 Matlab Simulink main page . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.18 The radar and the targets’ positions in the simulation . . . . . . . . . . . . 62
4.19 Phase noise block frequency response for the minimum and maximum phase
noise level 1. It is important to note that the x-axis is plotted in logarith-
mic scale in order to be able to depict the magnitude response transition.
Otherwise, in the linear scale, the magnitude drops very rapidly like a Dirac
function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.20 Phase spectrum for SNR = 10 dB and the phase noise level 1 = -80 dBc . 66
4.21 Phase spectrum for SNR = 5 dB and the phase noise level 1 = -80 dBc . . 66
4.22 The constellation of the complex signal at different SNRs and the phase
noise level 1 of -80 dBc: the color shading is getting stronger for the most
recent samples and the origin of the plane is denoted by the square. . . . . 67
4.23 Phase time trace for two different phase noise levels . . . . . . . . . . . . . 67
4.24 Fundamental frequency estimation error rates vs SNR and phase noise levels:
φn,1 is the phase noise level of 1. . . . . . . . . . . . . . . . . . . . . . . . . 68

5.1 Proposed signal processing chain . . . . . . . . . . . . . . . . . . . . . . . 70

5.2 Sitting tests setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

xi
5.3 Sleep tests setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.4 Vital signs waveform for hold-and-release-breath tests . . . . . . . . . . . . 75
5.5 Vital signs waveforms for the sleep tests on the back and right side: in each
figure, the original phase waveform together with the Hexoskin breathing
waveforms and the breathing waveform reconstructed from the optimum
filter are plotted. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.6 Vital signs waveforms for the sleep tests on the belly and left side: in each
figure, the original phase waveform together with the Hexoskin breathing
waveforms and the breathing waveform reconstructed from the optimum
filter are plotted. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.7 Heart waveforms for the sleep tests on the back and right side: in each
figure, the electrocardiographic signal (ECG) signal from Hexoskin vest and
the heart waveform reconstructed from the optimum filter are plotted. . . . 79
5.8 Heart waveforms for the sleep tests on the belly and left side: in each figure,
the ECG signal from Hexoskin vest and the heart waveform reconstructed
from the optimum filter are plotted. Note that the ECG signal when the
subject was sleeping on his belly is very noisy, which is due to improper
electrodes contact of the Hexoskin vest to the skin surface. The R peaks are
obvious to compare with the radar. . . . . . . . . . . . . . . . . . . . . . . 80
5.9 Radar BR comparison with Hexoskin (a), and radar HR comparison with
Hexoskin (b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

B.1 Propagation and collection subsystem . . . . . . . . . . . . . . . . . . . . . 101

B.2 Channel subsystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
B.3 Targets subsystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
B.4 Signal processing subsystem . . . . . . . . . . . . . . . . . . . . . . . . . . 102

xii
Abbreviations

ANC adaptive noise cancellation 8, 85

AR auto regressive 35

ARM advanced RISC machine 13

AWGN additive white Gaussian noise 10, 24

BR breathing rate 1, 2, 4, 11, 17, 49, 71, 82, 84, 85

BSS blind source separation 9

CFAR constant false alarm rate 100

CW continuous wave 3, 4, 6

DACM differentiate and cross-multiply 3

DDS direct digital synthesizer 2

DSP digital signal processor 13

ECG electrocardiographic signal xii, 76, 79, 80

FOV field of view 24

FTPR frequency-time phase regression 3

HMI hip motion intensity 17

HMUSIC harmonic multiple signal classification 35

xiii
HR heart rate 1–4, 11, 17, 71, 82, 84, 85

HRV heart rate variability 3, 4, 17, 74, 85

LNA low-noise power amplifier 20

LSE least square estimator 8

MUSIC multiple signal classification 9, 84

NF noise figure 23, 65, 85

NLLS non-linear least square estimation 29

PA power amplifier 20

PAPR peak-to-average-power ratio 20

PSD power spectral density 54

PWM pulse width modulation 51

RCS radar cross section 22, 53, 70, 100

RF radio frequency 1

RTT round-trip time 20

Rx receiver 13

SNR signal to noise ratio 5, 65

STFT short-time Fourier transform 43

SVD singular value decomposition 6

THD total harmonic distortion 63

TI Texas instruments 1, 9, 13, 85

Tx transmitter 13

xiv
UART universal Asynchronous Receiver/Transmitter 13, 93

ULA uniform linear array 62

UWB ultra-wideband 6

WSS wide sense stationary 22, 26, 36

xv
List of Notations and Symbols

Tr Chirp duration 19–21

Tc FMCW frame period 24
fb,max ADC sampling rate 24–26
Bs FMCW sweeping bandwidth 19, 22, 26
x Column vector xvi, 29, 35
fmax The frequency at the end of FMCW sweep 19
fmin The frequency at the start of FMCW sweep xvi, 19, 20, 22
tf Fast time 20, 21, 24
fc FMCW frame rate 24, 26, 27
λmax The wavelength corresponding to fmin 21, 26, 27
c Light speed 21, 22
xH Complex vector hermitian 36
A Matrix xvi, 29
k x k2 norm-2 squared of a column vector x 29
φn Phase noise 20, 21
A† Pseudo-inverse of a matrix A 29
ts Slow time 20, 21, 24, 26
xT Vector transpose 29

xvi
Chapter 1

Introduction

1.1 A big picture

As the analog and digital circuits technologies have been developed, semiconductor manu-
facturers have integrated both digital and analog parts on a single chip. This integration
not only reduces the size and the cost but also the power consumption, and makes it
attractive for energy-harvesting and portable systems. Indeed, the modern chip designs
equip sensors with built-in processors with which they can be connected to other devices
and chips. In fact, they are capable of being connected to others chips to make a network
of sensors.
One of the greatest companies with outstanding repute, Texas instruments (TI), has
developed modular mm-wave radars with seamless integration of the radio frequency (RF)
and analog circuits to digital circuits. The modularity and generality of the chips allows
advanced developers to mediocre users to adopt and configure them for their applications.
In this work, we will use them to detect and characterize the vital signs’ rates, namely
breathing rate (BR) and heart rate (HR). In general, the problem is developing a system
based on a mm-wave FMCW radar as a vibrometer. In the course of design and system
implementations, we have investigated all aspects by considering both analog and digital
aspects.

1
1.2 Literature survey
Monitoring vital signs of a human like heart rate and breathing rate is very critical for
saving lives. Individuals might suffer from a disease such as sleep apnea, or their lives have
been threatened by a natural disaster. In the former case, there is a need for continuous
surveillance of a patient in a room while in the latter case the problem is finding the victims
through obstacles like walls. In addition, if a patient’s skin is burned, it is not possible
to attach a device to the subject body to record and analyze their health-related signals.
One solution is using radars with the capability of sensing any environment dynamics. The
general principle of a typical radar is sending a specific signal then listening to its echo in
order to extract environmental features. This idea is not new as it has been used by bats
for million years ago (see [62]) in that they use acoustic waves. The acoustic waves could
not travel a long distance and they are easily interfered with any mechanical movements
and they need a medium for propagation. In contrast, electromagnetic (EM) waves can
propagate without a medium and can penetrate through obstacles and do not have the
limitations of acoustic waves.
Among popular applications of EM radars, biomedical sensing has gained increasing
attention for adopting them in remote sensing of vital signs such as HR, BR, blood oxygen
density, etc. For instance, radars can find HR and BR by detecting the chest wall move-
ment. This is greatly helpful for reducing the number of contact biomedical sensors for
monitoring of a patient over a long time. For example, to monitor sleep apnea, in [30] the
authors proposed a remote controlling system, which uses a 24 GHz radar.
As we are interested in using FMCW radar, it is important to note that there is no
comprehensive analysis for the Doppler accuracy of FMCW radars. But in [16], authors
investigated the range accuracy considering non-idealities in FMCW generation. This is
the first paper using direct digital synthesizer (DDS) for artificially adding non-linearities
and phase noise to evaluate their effects on the range detection. Using DDS helps to sweep
the amount of nonlinearity as well as the noise power. In fact, the phase noise power and
the FMCW ramp nonlinearity are represented by adding a Gaussian noise and a sinusoid
term (with much less frequency than the FMCW carrier frequency) in the phase of the
FMCW, respectively. Besides, they used short-time chirp-z transform for obtaining the
sub-micro meter range accuracy which is far less than the fundamental range resolution
of the radar. Due to importance of the phase noise on our system performance, we will
discuss the developed Matlab Simulink model to clarify the effect of the phase noise on the
beat signal phase as well as the end frequency estimator (see section 4.4).
In the generation of an FMCW signal, a sort of phase randomness appears, which
is called phase noise. While it is manifesting itself in the phase, it has impact on the

2
magnitude of the generated signal. In addition, the phase noise degrades the quality of
the parameter estimation in any radars particularly FMCW radars. In general, the phase
noise effect in the range detection of any radar kind is discussed in [19]. When the echo
signal is mixed with the transmitted signal at the receiver, the phase noise autocorrelation
becomes a function of transmitted phase noise autocorrelation. This phenomenon is called
the range correlation effect. In brief, [19] clarifies that the phase noise effect in the range
detection of continuous wave (CW) or FMCW radar systems increases as the range of the
subject increases. Fortunately, the phase noise due to the range correlation effect is very
low for our case because the subject is very close to the radar located at a few meters
distance. We will talk about this effect in section 3.2.
In [59], the authors focused on measuring the range error of an FMCW radar at 77
GHz. In fact, They used theories of previous works for analytical expression of the range
error rate lower bound. They showed that the range uncertainty increases by increasing
the range, which can be inferred heuristically.
The authors in [65] reconstructed the cardiac motion at different sides by claiming that
the cardiac and the respiration motions are isotropic meaning that the volume change of
lungs and heart can be detected from all directions. In contrast, they could not extract it at
the left/right sides as clean as the front or back sides. It is shown that by applying the pro-
posed polynomial curve fitting, most cardiac motion cycles have better performance than
the frequency-domain filtering technique. They compared the polynomial fitting method
with the traditional filtering technique, which has frequency-dependent phase shift, and
they found out that it is more suitable for time-domain analysis to remove the influence
of respiratory motion. Their method lacks enough evidence for convincing how the poly-
nomial interpretation of the breathing wave outweighs the traditional filtering, which are
seemed to be more practical. In [65], the radar sensor operated at a carrier frequency of
15 GHz with a 10-dBm power to detect the time-domain cardiac motion of the subject. In
addition, using the differentiate and cross-multiply (DACM) method for phase unwrapping
has been exploited.
Nosrati et al. in [54] used a CW radar at 2.4 GHz. Basically, for the heart signal
acquisition enhancement, the second derivative of the received phase is taken, which means
that the input phase signal is filtered by a filter with the frequency response of ω 2 . This filter
amplifies the higher frequencies pertaining to the HR. But, the final rate estimation of HR
will be biased with a higher weight on the lower frequencies. However, they claimed that
the accuracy is improved by almost 10% with the use of frequency-time phase regression
(FTPR) method instead of FFT.
For determining autonomic nerve activities, heart rate variability (HRV) is a helpful

3
sign, which can be characterized by different parameters [48]. HRV has two parts of
low (0.03-0.15 Hz)1 and high frequencies (0.15-0.45 Hz)2 . They used a 24 GHz radar
(JRC: NJR4265J1K) and achieved very accurate results of 93% with comparing to an
ECG recorder.
A CW radar equipped with a phased array is used to monitor two targets’ vital signs
in [52]. They designed the antenna array such that it has two beams at two angles.
They designed a dual-beam proof-of-concept radar with a 4-element linear array (for both
Tx/Rx). The two beams were at 25 and -35 degrees with respect to the antenna boresight.
In this way, the phase signature of two subjects located in the field of view of two different
beams can be separated. They tested the system for two subjects by detecting BR, but
HR detection is neglected.
In [18], a CW radar at 79 GHz is used together with a six-port for mixing the Rx/Tx
signals. Adopting a passive RF component instead of an active mixer reduces power and
cost. The outputs are four with different relative phase shifts. They have shown that this
radar can be used for the detection and characterization of the vibrating objects.
Similarly, in [15], the authors used the FMCW radar for the purpose of vital signs
detection. To choose the start frequency, they tested the FMCW radar at different start
frequencies with different bandwidths and different subject positions. At the end, they
came up with a great conclusion that the radar reliably senses vital signs for many cases,
however, they chose 9.6 GHz for the purpose of the demonstration and implementation by
considering other factors such as the low interference on the wireless local area networks.
They analyzed the magnitude and phase of the received echo signal and they showed
that the vital signs detection can be carried out on either phase or magnitude. This
result is against the discussion in [44] where the authors illustrated that the magnitude
analysis comes with higher order harmonics which degrades the vital signs’ rate detection
robustness, instead they exploited the phase extraction. Furthermore, [15] did not discuss
the linear phase trend that was apparent in the captured phase. This behavior seemed to be
the impact of the hardware non-linearities while they neglected with no more clarification.
A similar phase acquisition using FMCW radar at 80 GHz was presented in [66]. The
authors assumed that the distance of the target is known, therefore they did not concentrate
on designing an algorithm for finding the target ranges. In addition, the radar was tested
in front, back, left, and right sides of the subject when the target was sitting on a chair.
1
Low HRV frequency is related to both sympathetic and parasympathetic activities. The former regu-
lates the body’s unconscious actions and the later is responsible for stimulation of “rest-and-digest” that
occur when the body is at rest.
2
High HRV frequency contains parasympathetic activities

4
This experiment setup cannot be used for a practical clinical purpose when a patient is
supposed to be laying down on the bed with lots of background clutters. In addition, they
did not use phase unwrapping and we will show that is necessary in mm-wave frequencies.
It is also important to know the electromagnetic properties of the body to understand
how the wave is reflected. In [61], the authors investigated the main reason for reflecting an
EM wave from a body in microwave frequencies. In particular, there are two main reasons:
1) blood perfusion underneath of the skin surface leading to the skin impedance variation,
2) skin/body surface movement. By designing some experiments, they found out the body
displacement has a higher influence on reflecting the signal than the impedance change of
the skin surface. Their research was conducted in microwave frequencies at which a portion
of the wave can penetrate the body. While in mm-wave frequencies, as we consider, the
wave is totally reflected off the body surface.
The echo signal characteristics such as its polarization and power depend on the con-
stituent material of the target and its orientation. Consequently, in the literature numerous
studies are about echo signal characterization. There are two different purposes to inves-
tigate the EM modelling of a target: 1) exploiting the model for optimum design of the
radar hardware and signal processing algorithms for a particular application, 2) extracting
the properties by a radar for target classification. For instance, the EM characteristics of
the human body have been studied extensively in [32], [35], [8] from 10 Hz to 100 GHz.
In contrast, the authors in [49] proposed a single layer model of the torso while the
complete model is a multi-layer dielectric material. They developed the model based on
the fact that the reflection and transmission coefficients of the wave remain the same for
both models. In addition, for more simplification, they used only a portion of the torso for
simulating the vital signs to reduce the solution region in EM simulation. The vital sign
detection is also carried out in the EM simulator. This enabled them to compare their
model with the experiments. They used a stepped frequency continuous wave (SFCW)
radar operating at 2.5 GHz with 1 GHz of sweeping bandwidth in the experiments. Single-
subject and two-target experiments were conducted to show a good agreement between
the simulation based on the single-layer model and the practical measurements.
The use of other type of radars has been considered in the literature. For instance,
the authors in [40] proposed a time-varying filter to reduce noise in order to extract the
heart rate with an impulse radar. Initially, the algorithm finds the respiration rate. Then,
it adopts a time-varying filter after which the derivative of the received signal is taken to
increase the signal to noise ratio (SNR) for the estimation of the heart rate. Although
this paper considers how to increase SNR for detecting the heart rate, there was no thor-
ough description of the experiment equipment and the setup. Indeed, the accuracy of the

5
reference sensor has not been reported.
A ultra-wideband (UWB) radar is employed in [50] for individuals’ detection in a nat-
ural disaster such as under ruined buildings after an earthquake. Because the system is
wideband, it has low intrinsic SNR. Therefore, the victim’s breathing rate and his range
estimation is performed by taking singular value decomposition (SVD) of the so called
slow-fast time matrix. The SVD is used for the noise reduction by taking a portion of the
slow-fast time matrix containing most of the signal power. In this way, a subspace of the
matrix space is removed in which the power of the noise is dominant and it contributes to
decrease SNR. The singular values are scaling factors and determine the signal power cor-
responding to the space of that singular value. To determine the signal subspace, they set
a threshold on the singular values by which the signal space was found to be the subspace
spanned by the basis corresponding to the singular values greater than the threshold.
The selection between Heterdyne or Homodyne receivers along with DC and IQ im-
balance compensations are investigated in [29]. In Heterodyne receivers, there is an image
frequency band which falls in the desired band in baseband. The image frequency must be
eliminated by an RF BPF before mixing in the Heterodyne architecture. In contrast, in
Homodyne receiver the positive and negative frequencies will be overlapped on each other
in the baseband which increases the noise level (see section 2.1.1). However, this problem
can be solved by the use of IQ mixer. A complete comparison between Heterdyne and
Homodyne receivers are discussed in [29].
In CW radars with IQ mixers, the correlated signal, beat signal, has a constant phase,
which is a function of the distance to the target as well as the other constant delays between
the transmit and receive paths. If there is a vibrating object in front of the radar, the beat
signal will be a complex exponential with a frequency equal to the vibration frequency. But,
the constant phase determines the centre point of the arc around which the trajectory of
the complex points will be shaped in the real-imaginary plane3 . If we assume that the real
part of the complex is used for the vibration analysis, the small angle approximation is
only valid when the constant phase is around multiple of 2π. On the other hand, if it is
not, the phase of the real part will be a polynomial function of the phase, based on the
Taylor series expansion of cosine. Therefore, without knowing the distance of the target,
it will be hard to performe phase analysis by taking only real or imaginary parts of the
beat signal [29]. The problem is referred to as optimum/null point detection in CW radars
with real mixing. Furthermore, we shall address the real and IQ mixing in more details in
section 2.1.1.
3
The point cloud in the complex plain is not a complete circle in the microwave frequencies for the vital
signs detection as the phase is not rotating a complete circle. This is because the phase modulation is
occurred based on relative movement of the chest and the wavelength, which is very small in these bands.

6
 Among other radars, we chose FMCW for our study. Compared to the other alterna-
tives, mm-wave FMCW radars have a collection of advantages and features, which could
not be found in a single radar. Those are:

ˆ Being a mm-wave radar: The high attenuation in mm-wave frequencies provides a

high isolation between the co-located operating radars even if they are separated in
a few meters. Indeed, tiny displacements in mm are comparable to the wavelength
so they can be detected. This high sensitivity is required to detect the chest wall
movement, which is in mm order.

ˆ Discriminating range or localizing: The radar can distinguish the reflections from
different ranges; therefore, potentially it can be used for multi-subject vital signs
detection. This feature is recognized as the main advantage of an FMCW radar in
[66]. Indeed, high propagation attenuation reduces the possibility of having an echo
signal, which is bounced off multiple reflectors. Most probably, the echo signal is
reflected off a single object if the environment is not rich scattering. In that area,
the received signal at particular range experienced a line of sight wireless channel. In
contrast, CW radars suffer from multipath fading because they collect all reflections
from all objects at all visible ranges in a one sinusoid signal.

ˆ Being robust against thermal noise: FM signals are more robust against noise in
comparison to AM signals. Also, in FMCW radars the vital sign information is
encoded in the received phase similar to FM signals. Thus, FMCW radar is less
affected by the noise in comparison to impulse radars.

There are many efforts in literature for removing random body motions for vital signs
detection. In fact, the vital signs detection can be achieved when the body of the subject
is perfectly stationary, otherwise any random motion creates harmonics and distortions in
the signal such that detecting the vital signs becomes impossible. There are some solutions
proposed such as Changzhi et al. , in [43], proposed a two-Doppler radar system by which
they cancelled random body motions. They suggested the use of two radars placed on
the back and front sides of the body in order to have two different samples of the phase
modulation. By having them and considering the fat that the signs of the modulated phases
due to body motion are different, they can be combined such that the output contains only
the vital signs’ phase modulation. To implement the signal processing, they used phase
demodulation, which has advantages compared to the complex signal analysis [45]. To
exploit the benefits of the phase demodulation, the DC offset and IQ imbalances of the
complex signal should have been removed before moving to the phase domain. The DC

7
offset removal is a huge challenge ([43]) since it could not be distinguished with the DC
generated by the vital signs’ phase modulation (see section 3.3).
In addition to the body motion, self-motion of the radar with respect to the ground
coordinates should be removed. For example, if the radar is installed at the back of a driver
seat for health monitoring, the radar shakes due to bumps in the road. Fang et al. , in [70],
proposed dual-band Doppler radar, which uses the fundamental and harmonic frequencies
of the LO to have two Doppler radars. The Doppler radar operates at 2.4 and 4.8 GHz.
The fundamental frequency is used for vital signs detection while the other one is used for
the radar platform motion cancellation. More specifically, the harmonic Doppler radar has
Tx/Rx antennas pointed to a fixed reflectors to measure the radar motion. A single low-IF
receiver downconverts both fundamental and harmonic signals. By the two observed IF
signals, similar to [43], the radar motion can be cancelled. It is interesting to note that
the method used in [70] is not a simple subtraction or addition of the harmonic Doppler
signal; instead, they employed an adaptive noise cancellation (ANC) method, [67]. Here,
the reference signal is the harmonic response while the input signal is the fundamental
Doppler response. The ANC adopts an adaptive updating of filter coefficients by which
the desired error signal converges to zero. The zero error is equivalent to removing the
radar motion from the signal containing both radar and the vital signs modulations.
As mentioned before, one challenge in phase acquisition of the CW or FMCW radars
is the existence of a DC in the received signal. The reasons for such DC are mainly three
sources: 1) antenna coupling or RF cross-talk, 2) stationary object in a desired range, 3)
the target phase modulation. The last one should not be removed from the signal while
the other two should. There is no way to distinguish the first two DC sources from the
last one. Thus, we should resolve the issue either by removing the sources or accepting
the phase distortion. Authors in [38] investigated the DC cancellation methods proposed
in the previous works. There are different techniques for estimating the centre point of an
incomplete circle such as least square estimator (LSE) [22], linear demodulation [29, 56],
Yuen method [68], compressed sensing method [13]. Among them the compressed sensing
method represents more accurate results and we will describe it in section 3.3.
After the phase acquisition, the phase contains two periodic signals. In fact, the phase
is modulated by two vibrating objects. The aim is to find the fundamental frequency of
the periodic signals. This is a basic problem, which has been investigated extensively in
literature under the topic of f0 estimations, fundamental frequency estimation, and the
pitch detection, traditionally4 . The fundamental frequency estimator for a mixture of the
4
The terminology used in the speech analysis but we will use the term “fundamental frequency” esti-
mation.

8
periodic signals has a wide range of applications such as speech and image characteri-
zation and compression, individuals identifications etc. Extracting periodic signals from
a single channel noisy observation may be similar to the blind source separation (BSS)
problems; however, they differ in nature for having assumptions on the behaviour of the
sources, which are assumed to be periodic. When the original periodic source is mixed
with a noise, the mixed signal is not a periodic signal and it is called pseudo-periodic (or
quasi-periodic) signals. The problem boils down to the spectral analysis, which are divided
into two general categories of parametric and non-parametric methods. For example, au-
tocorrelation based techniques are parametric methods in which the autocorrelation of the
signal is used to determine the fundamental frequency. These methods are biased and
depend on the magnitude of the signal. Thus, if there are large fluctuations in the signal
magnitude, the method fails to find the true fundamental frequency [12]5 . Indeed, they
depend on a threshold, which has must be adaptively adjusted to maintain a low level of
error [66].
Furthermore, there are many parametric high resolution methods such as linear predic-
tion [21], subspace methods [25, 10], harmonic fitting [47], maximum liklihood [26], cepstral
method [51], and Bayesian estimation [20, 36]. Among subspace methods, multiple signal
classification (MUSIC) analysis exploits the signal subspace models to enhance, separate,
and estimate the periodic signal properties [23]. Christensen et al. , in [11], did not only
find the fundamental frequency, also they fully characterized the periodic signals. A com-
plete explanation and the mathematical derivation of the fundamental frequency and the
optimum filter design are given in section 3.5. We will use this method for the harmonic
analysis of the received radar phase for the vital rate estimations.
Authors in [10] developed an optimization problem in which the model order in [23]
and the fundamental frequency of a periodic signal will be obtained simultaneously. This
relies on a higher computational burden while it can be avoided by assuming a fixed model
order but loosing accuracy. We shall investigate each vital signs separately in chapter 5.
For our measurements, we used TI mm-wave radars in this work. One advantage is
that they leverage the benefits of IQ demodulators. They bring SNR improvement, noise
level estimation, intermodulation reduction, and simple range and phase demodulation.
We will discuss these in the next chapter.
5
This is why we will not use this method for the vital signs rate extraction.

9
1.3 Main contributions of the work
This work shows the feasibility and validity of using an FMCW radar to monitor vital
signs remotely. Also, we will investigate the theories required for fundamental frequency
estimations of two vibrating objects, in general. To achieve so, we used a mm-wave radar
from TI to analyze, develop, design and implement the signal processing. A summary of
the main contributions of the work are as follows:

1. A complete system model of a typical FMCW radar is presented including the an-
tenna coupling or RF cross-talk, phase noise, IQ imbalance, and stationary clutter.

2. The DC offset calibration is investigated and showed how it will affect the phase
quality.

3. The phase analysis is presented in which the phase wrapping is embedded and pro-
posed to uncover the true phase modulation. Indeed, the constraint for the proper
working of the algorithm is derived, which depends on the vibration frequency and
the magnitude of the vibrations.

4. An optimum filter design for the characterization and enhancement of the periodic
signals is included at the end of signal processing chain to find and estimate the vital
signs waveforms and their fundamental frequencies.

5. We designed experiments in which we had stationary objects to identify the linear

phase trend for the stationary targets.

6. We designed a two-pendulum system by which we were able to mimic the chest

wall movement due cardio-respiratory activities. The experiments helped to gain
intuitions for the vital sign detections.

7. We implemented the most complete FMCW radar system in Matlab using Simulink
model. The model gives a realistic sense of the system performance in the presence
of additive white Gaussian noise (AWGN), phase noise, IQ imbalance, stationary
clutters, etc.

8. Finally, the proposed signal processing is examined for vital signs detection. The
accuracy of the system compared to the other similar works using an FMCW radar.

10
1.4 Summary
In Chapter 2, we will introduce brief features and specifications of the used equipment for
the experiments. They are TI mm-wave radars and Hexoskin vest. We will also discuss
the benefits of the IQ mixing in a radar rather than using the real mixing.
In Chapter 3, the fundamental FMCW radar principles are discussed. This chapter will
substantiate the mathematical derivations required for the vital signs detection. Impor-
tant topics are DC compensation process, phase analysis of a vibrating object or a radial
movement, harmonic analysis, and real to imaginary phase transformation.
Chapter 4 verifies the proposed system with different experiments. They are stationary
tests, pendulum tests, and the Matlab simulation. We will discuss the results obtained by
the simulations, which are very close to the practical observations.
Chapter 5 contains experiments for the vital signs detection. Generally, two sets of
experiments were conducted. In the first set, a subject was sitting in front of the radar
facing the radar at a right angle. In the second set of experiments, a subject was sleeping
on a bed and the radar was mounted on the ceiling above the bed. The final results of BR
and HR accuracies are compared to the preceding works.
The work is concluded in Chapter 6 by wrapping up the novelties of the work and the
future extensions.

11
Chapter 2

Hardware and equipment

For this study, we use an off-the-shelf radar equipment by Texas Instruments for the
investigation and analysis of the signals for the final goal of vital signs monitoring. In the
subsequent sections, we will explain the general system parameters of each device used in
the experiments.

2.1 Texas Instruments mm-wave radars

In lieu of outstanding advances in the integrated circuit technologies, new ways of system
implementation and integration are offered by manufacturers. In fact, the traditional
complex systems containing different units and components are designed and developed
seamlessly both in hardware and software in a single chip. TI offers different mm-wave
FMCW radar solutions not only for a particular use but for a variety of applications.
Different applications range from proximity sensing, gesture recognition, advanced driving
assistance system (ADAS) for low range and middle range, occupancy detection, target
localization in 3D, vital signs detection to object identification and classification. There
are different demos and applications available in the project repository of TI accelerating
the design and developments. Some are discussed in Appendix A.
As mentioned in Chapter 1, mm-wave FMCW radar has benefits that cannot be found
in any single radar. Firstly, they are smaller and conveniently can be integrated to any
spaces such as the front bumper of a vehicle as a part of the parking assistance sensor, which
is the feature leveraged from small wavelengths in mm waves. In addition, the prominent
property of the FMCW radar is its capability in localization and range discrimination with

12
a single observation such that other types of sensors could not provide the target range
detection.
In addition, TI radars have multiple transceivers for exploiting multiple input multi-
ple output (MIMO) techniques in obtaining the angular information of the objects and
facilitating the 3D mapping of the environment. For instance, a flexible binary phase mod-
ulation (BPM) and time division multiplexing (TDM) can be used to increase the accuracy
of the azimuth and elevation angles’ estimations.
For our experiments, we chose a Texas Instruments (TI) mm-wave radar (AWR1443 [2]
or AWR1642 [4]) operating from 76-81 GHz. They have built-in digital signal processor
(DSP)s and advanced RISC machine (ARM) processor for post-processing. Although the
radar can run small applications stand-alone, we did not use it to run our processing
on the chip. In fact, after configuring the radar, the chirp samples are transferred over
the universal Asynchronous Receiver/Transmitter (UART) interface to a PC to performe
signal analysis in Matlab. To work with the radar, evaluation boards of AWR1443 [3] and
AWR1642 [5] were used.
AWR1443 has 3 transmitters and 4 receivers and AWR1642 has 2 transmitters and
4 receivers, but we use a single transmitter (Tx)/receiver (Rx) pair for the entire study.
According to this great difference between the two radars, the former can detect both
azimuth and elevation of the target, while the latter only can detect the azimuth angle. In
addition to mm-wave FMCW intrinsic features incorporated in these chips, they borrow
the assets of IQ demodulation, which enhance the system performance in comparison to
the real demodulation and we will discuss them in the following section.

2.1.1 Benefits of using IQ demodulator

The general architecture of an IQ demodulation and its counterpart, real demodulation,
are shown in figure 2.1. In the signal path, without loss of generality, the generator output
is a cosine-form signal with a phase φ(t). The signal goes to the transmit power amplifier
(PA) which is connected to the Tx antenna. The echo signal is collected by the Rx antenna
and the LNA amplifies the signal. The major difference between the two mixing methods
is that the power of the signal is divided between I and Q, while there is no power division
in real mixing. The following are the benefits of this structure.

13
 Figure 2.1: IQ demodulation vs real demodulation

SNR improvement

Although, the signal power splitting in the IQ mixing increases the contribution of the IF
component noises, the baseband noise due to RF components before mixing is smaller. In
fact, if the transmit power coupling between nearby Tx/Rx antennas and the RF cross-
talk in the RF circuit are the dominant sources of the noise, then IQ demodulation has
3 dB better SNR. Figure 2.2 explains this claim. Suppose that the reflected wave is the
superposition of the chirps with different delays associated with different objects in the
environment. Each delay is translated to a frequency shift of the chirp start frequency
and can be represented by the frequency components in figure 2.2. This concept is more
discussed in 3.1. By real mixing, the positive frequencies are shifted down to the left hand
side of DC and the negative frequencies are shifted to the positive side of the based band.
During the conversion, the noise level at the baseband is doubled due to overlapping noises
of the positive and negative frequencies. In contrast, in IQ mixing, only the positive (or
negative) portion of the spectrum is downcoverted keeping the noise level as it is in RF
–the IQ mixing shown in 2.1 downconverts the negative frequencies. Thus, there is at most
3dB more SNR in the IQ demodulation. Note that the relative magnitude of the peaks in
the frequency domain are drawn such that the farther objects have less magnitudes due to
more attenuation of the reflected wave –we assume that the radar cross sections (RCS) of
all targets are the same.

14
Figure 2.2: Real mixing: ∆f1 , ∆f2 , ∆f3 are corresponding to the downconverted frequencies
of f1 , f2 , f3 , respectively.

Noise level estimation

As mentioned before, while the IQ mixing results in an asymmetric spectrum at the based
band, this brings a benefit to having a portion of the spectrum1 containing only noise.
Therefore, it can be used for noise power estimation for the signal processing end.

intermodulation reduction

In real mixing, any image or interference out of band can be folded back to the band, which
degrades the signal quality at the digital input end. In practice, a portion of the transmit
power is coupled to the receiver either through RF corr-talk or antennas. Furthermore,
assume that a chirp reflection is received by a target. The delay of the coupled chirp
is less than the echo chirp, so the former has a closer start frequency (f1 in figure 2.3)
to the FMCW generator start frequency (fc in figure 2.3). By considering LNA as a
nonlinear component, it produces harmonics close to fc , which are falling in the same
band of the based band by means of real mixing. In figure 2.3, the lowest order in-band
intermodulations, which are ±2f1 ∓ f2 are illustrated with red peaks in the frequency.
The coupled chirp delay is very small, depending on the RF group delays between the
generator output and the receiver mixer, such that the intermodulation terms are so close
1
positive or negative frequencies depends on the mixing

15
Figure 2.3: Falling intermodulations of the LNA output in the based band: the dotted and
solid lines are for negative and positive frequencies, respectively.

to the desired echo chirps. In contrast, in the IQ demodulator, the intermodulation terms
are out of band and can easily be filtered.

Range and phase analysis

In the next chapter, the received complex signal of each range bin can be seen as the echo
signal in CW radars, but carrying the reflections only from one range. This means that
the same problem of optimum/null point detection [43] exists, which can be avoided by
using complex demodulation. Otherwise, in real demodulation the relationship between
the phase modulation and the basedband signal might not be linear2 .
2
For clarification, suppose that cos(ωr t + f (t) + θ0 ) is received by real demodulation where ωr , θ0
are the radian frequency and a constant phase corresponding to the range of a target, and f (t) is the
time-dependent delay consisting of the vibration waveform of the target. Thus, after moving it to DC it
becomes cos(f (t) + θ0 ). θ0 determines whether the cosine function can be linearly approximated to f (t)
or not. As a matter of fact, the relationship can be non-linear.

16
 Table 2.1: AWR1443 and AWR1642 specifications
Param. Tx min fc a Bs b max fb,max d max range phase
power /max Kc rangee reso- noisee
Rx lution
gain
Value 12 24/48 71-81 Up to 4 100 10 +100 3.75 -91
dBm dB GHz GHz M Hz/µs MHz m cm dB/Hz
a
Start frequency of the ramp
b
Sweeping bandwidth
c
Ramp slope
d
Maximum sampling frequency
e
For 77-81 GHz at 1 MHz offset

2.1.2 General parameters

Table 2.1 listed some of the important parameters of the radar, which are the same for
both AWR1443 and AWR1642. The phase noise is reported at 1 M Hz frequency offset,
which corresponds to 2.4 m range. The maximum range can be detected theoretically
by fb,max = 10 M Hz/µs is 150 m, which is fairly long such that the received power is a
restriction for the range detection. In fact, the max detectable range is limited by the
minimum required SNR 3.

2.2 Hexoskin vest

Figure 2.4: Hexoskin and BioRadio

The Hexoksin vest is a commercial product of Carre Technologies inc. [41]. They have
wearable smart textiles (figure 2.4) to monitor physiological activities including BR, HR,
HRV, tidal volume (vt ), minute ventilation, and hip motion intensity (HMI). The garment

17
can also record the position of a subject for 6 different postures: standing, sleeping on
stomache, back, left, and right sides. The accuracy of the HR and BR estimations for
different postures is evaluated by [63]. The authors claimed that the device measures BR
and HR for different body postures with 98% accuracy in comparison to the standard
laboratory measurement tools. We used this sensor as a reference for comparing our radar
results.

18
Chapter 3

System model and radar analysis

In any radar, the electromagnetic wave is sent into the environment containing various
objects. Then the echo of the wave is captured at a receiver. A simplified block diagram
of such a system is shown in Figure 3.1a in which both the transmitter and the receiver
are at the same location. Each chirp at the output of the FMCW generator is a sinusoid
signal whose frequency is swept from fmin to fmax (figure 3.1b). Here the frequency is
swept linearly with a positive slope of K and a duration of Tr implying that the sweeping
bandwidth is Bs = fmax − fmin = KTr . The received signal at the output port of the
receiver antenna is amplified and correlated with the transmit signal, which results in a
signal called beat signal. The beat signal contains information about the objects in the
scene. Particularly, the delay (td in figure 3.1b) in the reflected signal is translated to an
instantaneous frequency difference between the transmitted and the received chirps.

(a) Block diagram of an FMCW radar (b) Transmitted and received chirp sequences

Figure 3.1: Signal transmission and reception of FMCW chirps

19
 The power amplifier (PA) and the low-noise power amplifier (LNA) at the transmitter
and the receiver in Figure 3.1a are non-linear components. One feature of FMCW signals
is that they are constant envelope signals with the peak-to-average-power ratio (PAPR)
of zero dB, which lets the amplifiers work in their linear region. This adds time-varying
delays to the signal even if they are in their linear operation. The time varying delay is
much greater than Tr and it manifests itself as a very small Doppler shift in the baseband.
This Doppler shift is in slow-time as we will be seen later. In practice, the PA and the
LNA delays will diminish after a while when they become thermally stable.
Throughout the entire text, we use the terms “objects” and “targets” interchangeably
to imply that for the purpose of discussing the signal properties, it is not important whether
the signal is reflected off a desired object or a clutter.

3.1 FMCW radar fundamental principles

The complex chirp signal is [28]:

s(tf ) = At exp j(2πfmin tf + πKt2f + φn (tf )) ,

0 < tf < Tr , (3.1)

where At related to the transmit power, fmin is the start frequency (and λmax is the
corresponding wavelength), and tf is the fast time, which can be in [0, Tr ]. In fact, in
writing equations we should distinguish between the time that is within a chirp period or
the time that is within the physical displacement of the target. Also, φn is the phase noise
of the generator. we assume that for a single target when the thermal noise and any other
RF imperfections are ignored, the echo signal is a delayed version of the transmitted chirp
in (3.1):
 
2


r(tf , ts ) = Ar exp j 2πfmin (tf − td (ts )) + πK(tf − td (ts )) + φn (tf − td (tf )) , (3.2)

where td (tf ) is the round-trip time (RTT) delay of the wave which is a function of the the
slow time ts meaning that the delay changes over time are greater than Tr . Thus, again
we should emphasize that tf and ts must be distinguished.
We suppose that there is only a single small object positioned at a nominal distance of
R0 to the radar but it is moving from R0 , which results in a time-varying distance to the
radar. This time-varying delay should vary the order greater than Tr , if it is not, the chirp
duration should be smaller to meet the criterion. This assumption is very important will
be shown later. Let us denote this time-varying distance by R(ts ) = R0 +x(ts ), where x(ts )

20
is a function represents the distance variations. Furthermore, td (ts ) = 2R(ts )/c, where c is
the light speed throughout the whole text. As illustrated in Figure 3.1a, the beat signal
is the result of the received and transmit chirp product, i.e. y(tf , ts ) = s(tf ) r∗ (tf , ts ). By
plugging td (ts ) into (3.2) and doing some manipulation, y can be expressed as:

y(tf , ts ) = Ar At exp j 2πKtd (ts )tf + 2πfmin td (ts ) − πKt2d (ts )+

φn (tf ) − φn (tf − td (ts )) ) ,
   
4πK 4π 8πK
= Ar At exp j 2πfb tf + x(ts ) tf + − 2 R0 x(ts )
c λmax c

4πK
− 2 x2 (ts ) + θ0 + θ1 + φn (tf ) − φn (tf − td (ts )) , (3.3)
c
2KR0 fmin 4πR0 4πkR2
fb := , λmax := , θ0 := , θ1 := − 2 0 (3.4)
c c λmax c
Fourth and fifth terms in the exponential argument in (3.3) can be neglected because
typically, K/c2 is in the order of 10−4 (refer to the values in table 2.1). Thus, θ1 is
approximately equal to zero. The second term in (3.3) can be very small by considering
tf , ts values. We suppose a scenario in which the target has a radial velocity of 100 m/s. If
Tr = 10−5 , then ∆x = x(Tr )−x(0) is approximately equal to 1cm. Hence, during one chirp,
the coefficient in the second term would be around 10−3 and can be ignored. Similarly, if
the target is vibrating with fv frequency and a maximum physical displacement of 1 cm,
then the maximum ∆x during one chirp would be very small depending on the choice of
Tr 1 . Therefore,
  
4π
y(tf , ts ) ≈ Ar At exp j 2πfb tf + x(ts ) + θ0 + ∆φn (tf , ts ) (3.5)
λmax
4π
ψ(tf , ts ) = 2πfb tf + x(ts ) + θ0 + ∆φn (tf , ts ), (3.6)
λmax
∆φn (tf , ts ) = φn (tf ) − φn (tf − td (ts )),

In fact, ∆φn (tf , ts ) is the result of the range correlation [19] by which the effective phase
noise for the beat signal depends on the distance of the object. For instance, the closer
1
Suppose that x(ts ) = Av cos(2πfv ts ). If ff = 1/Tr , then ts = n/ff . Though, it can be shown that
the maximum ∆x = x(n + 1) − x(n) is Av (2πfv /ff ). Because this maximum is in the order of 10−5 , the
second term in the phase of the exponential function in (3.3) is almost zero.

21
the target is (td ≈ 0), the lower the phase range is. It is straightforward to obtain power
spectral density (PSD) of ∆φn (tf , ts ) as a function of the PSD of φn (tf , 0) by assuming
that the target is stationary and is located at R0 , and φn is wide sense stationary (WSS):
 
2 2πR0
S∆φ (f ) = 4Sφ (f ) sin f (3.7)
c

The PSD in 3.7 increases as the range of the target increases. According to the dependency
of the target range on the instantaneous frequency difference of the echo chirp and the
transmitted chirp (see figure 3.1b), the farther the object is, the more the frequency shift
is. In fact, it is true that the range correlation effect diminishes for longer ranges, but,
fortunately, in FMCW radars the phase noise power is less for the longer ranges since the
frequency offset to the start frequency (fmin ) is greater. In other words, even though if
there is no range correlation at a distant range,2 the phase noise power is much less than
the near objects. This also raises the need for minimizing the delay through the transmit
and receive paths to maintain the lowest possible phase noise effect at the output of the
mixing at the receiver.

3.2 Radar equation and received signal power

If the Pt is nominal transmit power and the target is at a distance of R, then the received
power is related to the transmit power of a radar by the following [17]:

Pt Gt Gr σλ2
Pr = (3.8)
(4π)3 R4

where Gt , and Gr are the transmit and the receive antenna gains, respectively. σ is the
radar cross section (RCS) of the target 3 . Also, λ is the wavelength of the travelling wave.
For physiological motion, the area under which the chest moves determines the RCS.
Equation (3.8) is known as the radar equation. We assume that the room temperature
is 300 Kelvin (or about 25 Celsius) and the transmit chirp has a sweeping bandwidth
of 4 GHz, then the noise power at the output terminal of the receiver antenna is Pn =
10 log10 (KT Bs ) ≈ −103 dB in which K is Boltzmann’s constant 4 . In addition, if the
2
because the frequency shift with respect to fmin is high
3
RCS is a hypothetical area required to intercept the power density at the target such that if the total
reflected power is scattered isotropically, the power density at the Rx is achieved.
4
K = 4.138 × 10−23 J/K. Do not confuse this K with the chirp slope.

22
receiver noise figure (NF)5 is NF dB and the minimum SNR required at the based band
is denoted by SN Rmin , then the minimum received signal power at the output port of Rx
antenna will be Pr,min = N F + Pn + SN Rmin in which NF is the noise figure and all the
variables are in dB. By using (3.8), which relates Pr to the target range, the maximum
range versus minimum required SNR at the based band can be obtained:
s
4 P G G σλ2
t t r
Rmax = , Pr,min = N F + Pn + SN Rmin (3.9)
Pr,min (4π)3

2.5
G = 10 dB
= 0.39 m2 G = 15 dB
2 NF = 16 dB
P = 12 dBm
t

1.5
(m)
max
R

0.5

0
0 5 10 15 20 25 30
SNRmin (dB)

Figure 3.2: Maximum range of a target versus minimum required SNR at the based band.
The gain of Tx/Rx antennas are the same as G.

In figure 3.2, Rmax is plotted versus minimum SNR with values annotated. In fact,
σ = 0.39m2 is an approximate value for a human as mentioned in [29]. The transmit
power of 12 dBm is the output power of AWR1642/AWR1443 chips (see table 2.1).

3.2.1 Complex signal analysis

Furthermore, ψ(tf , ts ) in (3.6) varies with x(ts ) relative to λmax . So, the phase variations
in the scale of the maximum wavelength can greatly change the beat signal phase. For
5
Noise figure is defined as the ratio between the input SNR to the output SNR of a circuit component.

23
example, a radar operating at 6 GHz is 10 times less sensitive in comparison to a 60
GHz radar. Thus, as a rule of thumb, the phase power for the same amount of physical
displacement is 20 dB more in mm-waves.
This equation is used to detect the range of a subject, R0 . To this end, an FFT is
applied over samples of a chirp to obtain the spectrum of the beat signal, which has peaks
corresponding to the subjects at different ranges. This FFT reveals range information
so it is called range FFT. Each range FFT bin represents a particular distance with an
associated phase similar to ψ(t).
The basedband signals y(tf , ts ), ψ(tf , ts ) (in (3.5) and (3.6)) are sampled in the fast
time by the sampling frequency of fb,max (see table 2.1), though tf = n/fb,max . In the
slow time, if each chirp is sent every Tc , then the sampling frequency will be fc = 1/Tc ,
though ts = m/fc . By replacing the discrete times in (3.5) and (3.6), the discrete form of
the signals will be obtained i.e. y[n, m], ψ[n, m].

Range detection

We suppose that within each chirp period the number of samples is N, then fb in 3.4 can
be estimated by means of peak searching in the FFT domain of y[n, m] over fast time
samples (with index n and FFT size of N). From that, an estimate of the target, R0 , will be
obtained. The following theorem proves that in an AWGN channel, fˆb is optimum in ML
sense when it is the maximum frequency component of y[n, m], when there is a single target
in the radar field of view (FOV). Also, note that Ar is (3.5) is not Rayleigh distributed
as usual wireless channels therein the assumption of the dense scattering environment is
a premise [37]. In contrast to them, here the channel is not a rich scattering environment
due to having limited number of reflectors at a distance R0 . Thus, Ar is not magnitude
of a zero mean complex Gaussian random variable. The following proves the optimality of
the range detection by simple peak frequency detection of y[n, m].

Theorem 3.2.1 (optimality of range detection) If y[n] and Y [m] are a discrete sig-
nal and its FFT ∀m, n = 0, · · · , M − 1, w[n] is an i.i.d. complex Gaussian random process
2π
with zero mean, and xp [n] is in {ej M pn  p = 0, · · · , M − 1}, such that:

y[n] = xp [n] + w[n]

then, the maximum estimation of p is:

 
p̂ = argmax Y [m] (3.10)
 
m

24
 Proof: The problem is to find a xi [n] such that the following will be satisfied in an
interval of M samples:   
max pr y[n]xi [n] (3.11)

i

in which pr is the conditional probability. (3.11) is maximum likelihood (ML) criterion and
it is optimum when all possible xi [n] can happen with equal probability and equal power.
ItEven though can be shown that (3.11) is equivalent to the minimum Euclidean distance
criterion between y[n] and xi [n]. Then (3.11) is equal to the following:
 
M −1 
X
max  y[n]x∗i [n] = max R (y[n], xi [n]) (3.12)

i  n=0  i

and R is the cross-correlation function between the two sequences in its argument. It is
enough to show that the following is true:
 2πin

argmax Y T ei  = argmax R y[n], e−j M
 
(3.13)
i i

where Y is a vector containing frequency samples of y[n] from discrete-time Fourier trans-
form (DFT). ei is a vector having only one at the i’th position. We start to expand the
inner argument of the left-hand side of the equation:
 
T j 2πn i
Y ei = DF T y[n] ~ e M (3.14)
−1
M
!
2πi
X
= DF T y[l]ej M (n−l) (3.15)
l=0
−1
M
"M −1 #
1 X X
j 2πi 2π
= y[l]e M
(n−l)
e−j M nk (3.16)
M n=0 l=0
M −1 M −1
X 2π
X 1 −j 2πn (i−k)
= y[l]e−j M il e M (3.17)
l=0 n=0
M

~ is a circular convolution is used to convert frequency multiplication to time-domain

convolution. In (3.16), k is the index of FFT and l is the index of convolution. Af-
ter interchanging the summation orders in (3.16), (3.17) is obtained in which the second
summation is δ(i − k). By replacing it in (3.17), the theorem is proved.

By the Nyquist sampling theorem, fb,max /2 should be greater than or equal to fb to be able
to detect fb (or the range of the target). Therefore, the maximum unambiguous detectable

25
 Figure 3.3: Radial velocity with respect to the radar

range will be Rmax = fb,max /2K, which is followed from 3.4. Also, the range resolution,
which is the minimum resolvable distances, is Rmin = fb c/2N K = c/2Bs , where Bs is the
sweeping bandwidth.

Radial displacement

After applying FFT on y[n, m] over n indexes and removing the fb frequency, the resulting
signal is the slow time variation of y i.e. ỹ[m] = FFTn (y[n, m]). In addition, because φn [n]
is WSS, so without loss of generality, we set n = 0. Thus,
  
4π
ỹ[m] = Ar At exp j x[m] + θ0 + ∆φn [m] (3.18)
λmax
4π
ψ̃[m] = x[m] + θ0 + ∆φn [m] (3.19)
λmax
where ∆φn [m] = ∆φn [0, m]. Two cases are of interest for x[m] function. Here, we investi-
gate when the target has a radial velocity of v with respect to the radar (Figure 3.3) and
when it is modelling the vibration of the target. In Figure 3.3, the target is moving towards
or away from the radar determining the sign of v in x[m] = vts = mv/fc . Hence, ỹ[m] is a
periodic signal with ωr = 4πv/λmax fc , which is the normalized radian frequency. So, again
by taking the second FFT across the detected range bin samples, an estimate of ωr can be
obtained. In fact, fc should be high enough such that |ωr | ≤ π. It is interesting to note
that in ωr , the velocity is measured with respect to λmax implying that the sensitivity to
the minimum detectable velocity of the radar is managed by its operating frequency. For
example, at 77 GHz and v = 1mm/s, then ωr = 3.22 Rad/s if fc = 1 Hz. In contrast, the
maximum velocity should be fc /(4λmax ) such that the Nyquist condition is satisfied.

26
Vibration displacement

If the target is vibrating around R0 instead of moving away or towards the radar, then
x[m] = Av cos(ωv ts ). By substituting it to 3.18, then:
  
4π
ỹ[m] = Ar At exp j Av cos(ωv /fc ) + θ0 + ∆φn [m] (3.20)
λmax
∞  
j(θ0 +∆φn [m])
X
l Av
= Ar At e j Jl 4π ejl ωv (3.21)
l=−∞
λ max

the second equality is the result of the Fourier Series (FS) expansion of the right hand
side equation in 3.20. Jl is the l’th order Bessel function of the first kind. Though, in the
frequency domain, the vibration produces many harmonics make the spectral estimation
of the ωv inaccurate where Jl is not necessarily decreasing as l increases. So, for instance,
if there is a maximum peak at fv1 it might be for a harmonic l > 1 rather than the
fundamental vibration frequency. The problem of estimating the fundamental vibration
frequency is more discussed in 3.4.
In fact, for vital signs detection application, x[m] can be a function models the chest
wall displacement of a subject. As it is vibrating due to the heartbeat and exhalation or
inhalation, it is a periodic function [28]. In that case, two vibrating objects at the same
location modulating ψ̃[m]. Thus the phase of the complex signal in 3.18 contains two
periodic signals. As a result, the FS expansion of each periodic signal will be multiplied
producing many harmonics. So practically, estimating the two vibration frequencies be-
comes infeasible. In next, we will provide a brief explanation about the two vibrating cases
to highlight some important points that are needed to be considered during analyzing the
vital signs.
To monitor respiration and heart rates, the chest wall can be modeled as two vibrating
objects with different frequencies and different magnitudes. Suppose that fb , fh , Ab , Ah are
BR, breathing amplitude, HR, and heart amplitude, then the phase of ỹ[m] will be:
   
4π fb 4π fh
ψ̃[m] = θ0 + Ab cos 2π m + Ah cos 2π m + ∆φn [m] (3.22)
λmax fc λmax fc

[43] showed that ỹ[m] with the phase in 3.22 will produce so many harmonics as well
as DC terms in real and imaginary parts. We will investigate the DC problem in phase
analysis, the DC terms due to vibrations do not have a detrimental effect on extracting
the phase of ỹ[m] because they have resulted in by the desired phase modulation of the
vital signs.

27
3.3 DC compensation process
The received signal in the form of (3.18) does not have DC term. However, in practice,
there are reasons for having DC terms in the reflections. As [29] mentioned, there are four
main sources of generating the DC:

ˆ Leakage between Tx and Rx antennas [70]

ˆ RF cross-talk between Tx and Rx paths [70]

ˆ Stationary clutters [29]

ˆ Phase modulation by a vibrating object [43]

The result of the DC in the complex plane is that the signal constellation is shifted from
the origin to a new point x (x = [dcr dci ]T ). In Figure 3.4, a constellation shift is apparent
in the point cloud before DC compensation. If there are any DC terms rather than the
phase modulation, then they should be removed or minimized before the phase analysis
otherwise in the phase domain there will be so many harmonics. In contrast, the phase
modulation itself results in a DC and it must not be removed since it is a part of the phase
modulation. Therefore, there is a challenge to keep the desired DC term while removing
unwanted DCs. Practically, this can not be achieved while there is no mean to distinguish
the two sources of the CD generators. If the unwanted DC term is di + jdq , the phase is
corrupted and it is shown by the following manipulations:
 
Q[m] + dq
ψ̃DC [m] = arctan (3.23)
I[m] + di
in which Q[m], and I[m] are the imaginary and real parts of the DC-free signal in (3.18).
(3.23) can be expressed as a Taylor series expansion around the desired point i.e.:
  Q[m]
Q[m] 1 I[m] 2 2


ψ̃DC [m] ≈ arctan + 2 ∆ψ[m] −  2 ∆ψ [m] + o ∆ψ [m] ,
I[m]
 2 
Q[m] 
1+ I[m] 1+ Q[m]
I[m]

(3.24)
Q[m] + dq Q[m]
∆ψ[m] := − , (3.25)
I[m] + di I[m]
The first term in the right hand side of (3.24) is the desired phase while the other terms
are the harmonics and intermodulations caused by the DC. It is straightforward to replace

28
Q[m] and I[m] by sin ψ̃[m] and cos ψ̃[m], respectively, and use the trigonometric identities
to obtain an equivalent equation like below:
1 
ψ̃DC [m] ≈ ψ̃[m] + 1 + cos(2ψ̃[m]) ∆ψ[m]+
 2 
1 1
sin(4ψ̃[m]) + sin(2ψ̃[m]) ∆ψ 2 [m] + o ∆ψ 2 [m]


(3.26)
8 4
This equation reveals the fact that the phase has similar harmonics as in (3.21).
The center and the radius of the cloud in Figure 3.4 can be estimated based on a non-
linear least square estimation (NLLS). Let us denote each complex point as a column vector
whose first and second elements are real and imaginary parts of the point, respectively.
Furthermore, assume ai , x, r are the i’th point, hypothetical centre point and the radius,
respectively. For finding an estimate of the centre point and the radius we solve the
following NLLS problem6 :
N
X
2
P = min k x − ai k2 −r2
x,r
i=1
N
X
2
= min k x k2 − 2aTi x+ k ai k2 −r2 (3.27)
x,r
i=1

N is the number of samples. The last summation resembles the second norm of a vector.
By rearranging the inner bracket terms and putting unknowns in a vector, y, and known
variables in a vector b the following equation can be obtained:
P = min k Ay − b k2 (3.28)
y

y = [R xT ]T , R =k x k2 −r2 , and A, b are:

   
1 −2aT1 − k a1 k 2
1 −2aT   − k a2 k 2 
2
A =  .. ..  , b =  (3.29)
  
.. 
. .   . 
T 2
1 −2aN − k aN k

A is a full column rank matrix for N > 3, thus has a left inverse of A† and the optimum
solution is:
y ∗ = A† b = (AT A)−1 AT b (3.30)
6
This method is known as compressed sensing method for estimating and tracking the centre and the
radius of an arc [13].

29
 150
Before DC compensation
Origin
100 Shifted origin
After DC compensation

Q 50

-50

-100

-150
-150 -100 -50 0 50 100 150
I

Figure 3.4: Correction of the received complex signal.

The optimum solution of P in (3.27) is equal to the optimum solution in (3.28) when
r =k x k2 −R > 0. By contradiction, if r < 0 and y ∗ = [R xT ]T is optimum the following
shows y ∗ is not optimum:

0 ≤k x − ai k2 =k x k2 −2aTi x+ k ai k2
< R − 2aTi x+ k ai k2 , ∀i (3.31)

therefore each term of the summation in (3.27) with y ∗ is greater than when y = [k
x k2 xT ]T implying that r has to be zero. Figure 3.4 is obtained for the actual vital
signs’ waveforms. The mm sensitivity makes the complex signal trajectory in Figure 3.4 a
complete circle. This is the result of the phase dependency on the relative chest movement
to the wavelength causing a great phase change only for displacements in the range of mm.

3.4 Phase analysis

In section 3.2.1, we represented that the complex received signal modulated with two
vibrating objects at the same range has too many harmonic components. Instead, for vital
signs application, we recommend to use phase analysis (or traditionally known as arctan
demodulation) in order to reduce higher order harmonics of the breathing on the heartbeat
band.

30
 Suppose that the received basedband complex signal has zero phase meaning that it is
real, a. If a complex noise, n(t) = nc (t) + jns (t), is added to the signal, then it deviates
the actual phase of the received complex sample by ∆θ:
y(t) = a + nc (t) + jns (t) = a + n(t) = a + Bn ejθn (t))
   
−1 ns (t) −1 ns (t)
θy (t) = tan ≈ tan (3.32)
a + nc (t) a
ns (t)
θy (t) ≈ (3.33)
a
3.32 is approximated when nc (t) is much smaller than a, though the noise shifts the phase
by:
Bn
θy (t) ≈ sin(θn (t)) (3.34)
a
Bn is the magnitude of the noise. The power of the θy (t) is the squared expectation of it
and since it is zero mean, then:
1 σn2
E(∆θ2 ) ≈ (3.35)
2 a2
where σn2 is the noise power, and σn2 /a2 is the inverse of SNR. Therefore, as SNR increases,
the phase deviation due to the noise decreases.
In practice, the phase is calculated by tan−1 function which outputs phases in [−π, π].
If ψ[m] in equation (3.5) is sampled with an appropriate sampling time of Tc then it is
possible to maintain the phase difference between two consecutive samples less than π.
Equivalently, x(ts ) (in (3.3)) must not change more than λmax /4 within Tc period. If this
assumption is satisfied, any phase change greater than π indicates that the phase should
be corrected by adding or subtracting 2π. This process is called phase unwrapping. In
other words, there should be at least three phase samples in [−π, π] to be able to recover
the phase, which is modulated by a physical displacement.
We should do the phase unwrapping across the phases on a range bin where the target
is sat. Let us denote the phase sequence as φn for n = 1, · · · , M , namely M is the size of
the vibration FFT. For every pair of (φn ,φn+1 ), we know that |φn+1 − φn | should not be
greater than π. If it is then a phase correction must be applied. The phase correction is
accomplished as bellow:

ˆ φn+1 − φn > π: means that φn+1 > 0 and φn < 0. As illustrated in Figure (3.5,
left), it seems the target followed the path shown in the red color just in Tc period.
However, with our assumption it should pass the green path, which is shorter and it
is less than π. To achieve so, 2π must be subtracted from φn+1 .

31
Figure 3.5: Phase jump cases: phase difference of greater than π (left), phase difference of
less than −π (right). The red lines are the unacceptable paths for phase change while the
green lines are the acceptable paths.

ˆ φn+1 − φn < −π: means that φn+1 < 0 and φn > 0. Similarly, in Figure (3.5, right) it
is shown that the target followed the long path with the red color while the correct
one is the shorter one. To correct, this time 2π must be added to φn+1 .

The whole process must be done in a sequence starting with n=1 and ending with
n=M-1. At each iteration, φn+1 is only updated. The algorithm 1 explains the process.

Algorithm 1 Phase unwrapping procedure

r is the desired detected range index r desired = R(:, r) (R is the range-slow time matrix
and Matlab syntax is used to select the r’th column) φ = tan−1 (r desired ) for n=1:M-1
do
if φn+1 − φn > π then
φn+1 = φn+1 − 2π;
else if φn+1 − φn < −π then
φn+1 = φn+1 + 2π
else
Do nothing
end

Figure 3.6a shows an example of a phase function before wrapping, after wrapping
and after applying the unwrapping process in algorithm 1. In this example, the phase
magnitude is growing exponentially with fixed sampling frequency. The growing magnitude
increases the phase difference between consecutive samples gradually. Thus, it can be seen
that the phase unwrapping process fails when there is less than 2 samples in [−π, π]. One

32
way to correct unwrapping for a sparse sampled phase, which may happen when the object
moves faster than it should, is to upsample the complex signal. By upsampling, we add
more samples between I and Q samples. Thus, there will be more phase samples helping
to have more change to unwrap the phase properly. As an illustration, in figure 3.6b, the
phase is upsampled with a factor of 4 so that it is possible to recover the phase even when
the phase without upsampling could not be properly unwrapped.

33
 20

15
Original
10 Wrapped
pi
-pi
5 3pi
Phase (Rad)

-3pi
0 5pi
-5pi
Unwrapped
-5 Upsampled &
unrwapped
-10

-15

-20
0 0.5 1 1.5 2 2.5 3 3.5 4
Time (s)

(a) Before t = 0.25s, there is no need for phase unwrapping. Also, the phase correctly unwrapped before t = 2 s
either for the phase with the original sampling rate or for the upsampled phase. But the upsampled phase is
correctly unwrapped until t = 2.18s.

20
Original
Wrapped
15 pi
-pi
3pi
10
-3pi
5pi
5 -5pi
Phase (Rad)

Unwrapped
Upsampled & unrwapped
0

-5

-10

-15

-20
1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8
Time (s)

(b) Zoomed-in figure 3.6a

Figure 3.6: Phase unwrapping test

34
3.5 Harmonic analysis
After unwrapping ψ̃[m] in (3.19), a quasi-periodic signal will be obtained. The signal could
not be an exact periodic signal because of additive noise (see (3.35)), phase noise, and im-
perfect DC compensation. For vital signs detection, the received phase is a superposition
of two periodic signals. Hence, the problem turns out to be fundamental frequency es-
timation from a mixture of periodic signals i.e. multi-pitch processing. Commonly the
problem is addressed in many fields such as speech recognition and compression, image
compression and feature extraction. In literature, there are many different approaches to
estimate the fundamental frequencies of a signal [27]. Among them, we choose harmonic
multiple signal classification (HMUSIC) methods presented in [24], [25] and [10]. We use
their final results, so for more details refer to the papers. Here we will describe the problem
in vital signs monitoring terminology as well as additional discussions on the noise.
Let us assume that we have K quasi-periodic signals in the observed signal x(t) corre-
sponding to the breathing and heart signals. Each quasi-periodic signal can be expressed
as follows:
Lk
X
xk [n] = ak,l ejωk ln + ek [n] (3.36)
l=1
ak,l is the complex coefficient of the k’th signal component and l’th harmonic order. Corre-
spondingly, ωk is the fundamental radian frequency of the k’th component. Each periodic
signal has its own noise, ek (n), which is uncorrelated (in the case of complex Gaussian
noise they are also independent). Lk is the model order of the k’th component. Also note
that the complex coefficients are not varying with time and they are ak,l = Ak,l ejφk,l with
Ak,l is a positive constant and φk,l is uniformly distributed in [−π, π]. The received signal
is a superposition of individual signal components:
K
X Lk
K X
X
x[n] = xk [n] = ak,l ejωk ln + e[n] (3.37)
k=1 k=1 l=1

where e[n] is the aggregated noise of all components. For the vital signs signal, there
are only two components of breathing and heart signals, so K = 2. By taking M signal
samples, we define a vector x[n] = [x[n], x[n − 1], · · · , x[n − M + 1]]T . An important note
is that this method is a prediction method based on M previous samples. M should be
at least equal to the longest period in (3.37) corresponding to the smallest fundamental
frequency. It is interesting to note that M is the order model in auto regressive (AR) for
(3.37) but the model here is used to estimate ωk and ak,l coefficients instead of predicting
the signal at future. Indeed, because we are assuming that the complex coefficients are

35
WSS, thus the signal should be WSS otherwise the model is not held. WSS assumption
is violated when there is a sudden short-time event in the signal such as a change in the
average or the presence of an interference. By rearranging and putting variables in (3.36)
to matrix and vectors, xk [n] can be expressed as:

xk [n] = Ak [n]ak + ek [n], (3.38)

 
ejωk n ej2ωk n ··· ejωk Lk n
 ejωk (n−1) ej2ωk (n−1)
··· e jωk Lk (n−1) 
Ak [n] = 
 
.. .. .. 
 . . ··· . 
jωk (n−M +1) j2ωk (n−M +1) jωk Lk (n−M +1)
e e ··· e

0
 
1 1 ··· 1 ejωk n
 e−jωk e−j2ωk ··· e−jωk Lk 

ej2ωk n

 
=

.. .. .. 
 . ..

 . . ··· .  

e−jωk (M −1) e−j2ωk (M −1) · · · e−jωk Lk (M −1) 0 ejωk Lk n
(3.39)
:= Zk Dn (3.40)
:= Zk [n] (3.41)

(3.39) is derive by looking at each column of Ak so that the ith column is multiplied to
ejωk in . We split the matrix Ak into two matrixes of one independent of time (Zk ) and
the other is a diagonal matrix containing time information (Dn ). Consider that Zk is a
Vondermonde matrix which is a full column rank matrix when M ≥ Lk (has more rows
than columns). In (3.40), matrix Dn only depends on n and by plugging it to (3.38), either
the matrix Zk can be seen as a time varying matrix ((3.41)) or the vector ak :

xk [n] = Zk [n]ak + ek [n] = Zk ak [n] + ek [n] (3.42)

x[n] is the summation of xk [n] and it has an correlation matrix of R:

K
X
H
R = E(x[n]x [n]) = Zk Pk ZkH + Q (3.43)
k=1

in which Pk is a diagonal matrix if ak has uncorrelated coefficients for different l with diago-
nal elements of A2k,l . Q is the accumulation of the source noise powers. An estimation of the
−1
correlation matrix of R can be found by sample averaging: R̂ = 1/G m=M (E(x[n]xH [n])
P
N −1

36
and G = N − M + 1 and N is the total number of samples over which the average is com-
puted7 . R̂ is invertible when it is full rank and it happens when M < N/2 + 1 because R̂
will be the summation of at least M rank 1 matrixes.

3.5.1 Optimum filter design

In order to find an estimate of ak and ωk , one can design a filter such that its output is
the closest signal to the noiseless component in (3.36). The idea is similar to designing a
matched filter, but here the output power of the filter is maximum when the true values
of ak , ωk are found. Assume, hk [n] = [hk [0], hk [1], · · · , hk [M − 1]]H are the complex filter
coefficients, then we want to minimize the numerical average error between the output
filter samples and the noiseless signal component i.e.:
N −1
1 X  2
P = min y k [n] − ŷ k [n] (3.44)

ak ,ωk ,hk G m=M −1
N −1
1 X  H T
2
= min h x[n] − a w [n] (3.45)

 k k k
G m=M −1

ak ,ωk ,hk

In (3.45), wk [n] := diag(Dn ). Consider that the objective in (3.45) is minimizing the
average error between the filter output and the ideal noiseless periodic signal. Also, we
have assumed that Lk is known while there are methods for estimating the model order (see
[10]). The problem in (3.45) is quadric
P −1in both ak and kk . In addition, it has a bounded
optimum solution when Wk := 1/G N H
m=M −1 w k [n]w k [n] is positive semi-definite. In fact,
it is straightforward to show that Wk is positive semi-definite 8 .
Assume that ωk is known. By taking derivative of (3.45) with respect to ak , the
optimum harmonic coefficients will be obtained as:

aˆk = Wk−1 Gk hk (3.46)

P −1
and Gk := 1/G N H
m=M −1 w k [n]x [n]. By replacing (3.46) to (3.45) it will turn to a new
problem for hk , which is quadratic and has a trivial solution of zero. Thus, a new condition
on hk will be added such that the filter impulse response has 1’s on the harmonic frequencies
7
Because is is assumed that R̂ is stationary, then N should be long enough to have a good estimation of
R. Practically, it should be at least 5 periods of the periodic component having the smallest fundamental
frequency.
8
If U ∈ Rn×n and U = vv H , then ∀x ∈ Rn , xH U x =k v H x k2 ≥ 0.

37
Algorithm 2 Finding ĥk , âk , ω̂k
Set N,M,L // based on prior knowledge about the signal
cc = 1 // counter
for ∀ωk ∈ Ω do
Compute: Gk , Wk , Q̂k
Compute: Hmat (cc, :) = hˆk // The matrix indexing syntax is the Matlab style.
The right hand side is from equation (3.47).
Amat (cc, :) = aˆk // from equation (3.46)
p̂k (cc) = hHk R̂hk
cc = cc + 1
end
Find î = index of max(p̂k )
ĥk = Hmat (î, :)
âk = Amat (î, :)
ω̂k = i’th member of Ω

of ωk l for l = 1, · · · , Lk . After solving the new problem, optimum filter coefficients can be
calculated by:
ĥk = Q̂−1 H −1 −1
k (Zk Q̂k Zk ) 1 (3.47)
1 is a column vector of all ones, and Q̂k is the residual noise correlation matrix remained
−1
after estimating ak . In other words, Q̂k := R̂ − GHk Wk Gk .
For estimating ωk , a brute-force method is adopted by testing all candidate ωk ’s so
that the filter output power is maximized i.e.:
ω̂k = argmax hH
k R̂hk (3.48)
ωk ∈Ω

and Ω is the set of all possible ωk ’s. For instance, breathing rate is in [0.1, 0.6] Hz. One can
search within this frequency range by a frequency resolution of ∆f to obtain a profile for
the filter output power versus all frequencies. In ideal, ω̂k is corresponding to the profile
peak. In summary, the complete algorithm for finding the filter, harmonic coefficients, and
fundamental frequency is shown in algorithm 2.

Residual noise power in successive signal component elimination

As we will see in the vital signs chapter, the breathing signal produces large interference
on the heart band leading to a poor estimation of the heart rate. One suggestion might

38
be using (3.47) and (3.46) to construct a noise-free breathing signal, then subtract it from
the signal to enhance the heart signal. This must be done with the noise and distortion
considerations.
Suppose that the first component in (3.37), which is for breathing signal, is perfectly
estimated with zero error, thus y[n] = hH
1 x[n]. Therefore, the difference x[n] − y[n] is a
new observation, which can be used for heat signal characterization by applying the same
procedure in algorithm 2. So,
L1
X L2
X
d[n] = x[n] − y[n] = jω1 ln
a1,l e a1,l ejω2 ln + e[ n] − hH
1 x[n] (3.49)
l=1 l=1
= wT1 a1 + wT2 a2 − hH n H n H
1 Z1 D1 a1 − h1 Z2 D2 a2 + e[n] − h1 e[n] (3.50)

the third term in (3.50) will cancel the first term and after rearranging terms, the difference
signal will be:
H
d[n] = (wT2 − hH n
1 Z2 D2 )a2 + h̃1 e[n] (3.51)
H
where h̃1 = [1 − h[0], h[1], h[2], · · · , h[M − 1]]H , in which all the entries are the same as
h1 except the first one. So, the new noise power is:

P̃n = σn2 k h̃1 k2 (3.52)

Because there is no restriction on the k h1 k during optimum filter design, so the noise
power could be more when a component is subtracted from the original signal. Also note
that the second term in (3.51) corrupts the heart signal.

3.5.2 Harmonic processing complexity

The complexity of algorithm 2 depends on the calculation of wk , Wk , Q̂k , ĥk , R̂, and ak .
But, the complexity order has an upper-bound denoted by O [14]9 . The overall calculation
of the mentioned variables is O(M 3 ) and M is defined as in algorithm 2. There is an
iteration on all ωk ∈ Ω and if |Ω| is the cardinality of the candidate frequencies, then the
total time complexity of the algorithm is O(|Ω|M 3 ). For instance, if breathing frequency
is divided in 100 segments with M = 200 when the sampling frequency is 20 Hz, then
800, 000, 000 operations are required to get the parameters.
9
See chapter 3.

39
3.5.3 Real phase to complex phase
The phase signal in (3.19) is a real signal but the model in (3.36) is complex. In fact,
every real signal can be converted to an analytical downsampled complex signal [7]. The
complex signal has the same real value as the original real signal. There is another method
for making a complex signal out of a real signal such that they have the same real parts.
The idea comes from the fact the one can set negative frequencies of the real signal to
zero without loss of information. In fact, the real data has a symmetric frequency content
so that having half of the frequencies is enough for representing the signal. To set the
negative frequencies of a real signal to zero and maintaining the same real parts, Hilbert
transform of the signal is required as follows:
1
y[n] = √ (x[n] + j x̂[n]) (3.53)
2
in which y is the complex signal and x̂ is the Hilbert transform of x. But y in 3.53 is
not analytical while its counterpart for the continuous signals is [7]. To have an analytical
signal, the real and imaginary parts of y should be orthogonal, which leads to the following
conversion written in Matlab syntax.

Listing 3.1: Matlab code for real to complex conversion

N = length ( x ) ; X = fft (x , N ) ;
y =( l /2) * ifft ( [ X (1) + X ( N /2+1) ;2+ X (2: N /2) ] , N /2 ) ; % By downsampling
factor of 2 , the two ends of the spectrum will be overlapped .

3.6 Summary
In this chapter, we started the discussion with fundamental principles of FMCW radar. The
mathematical derivation for obtaining target’s range, velocity, and vibration frequency was
presented. Based on the parameters defined in Table 3.1, the final equations for estimating
each parameter are listed in Table 3.1.
In practice, before phase analysis, the DC of the received complex beat signal should be
removed. After removing DC, the phase is calculated and unwrapped based on algorithm
1. Furthermore, the real phase should be converted to an analytical complex signal prior to
fundamental frequency estimation. The algorithm 2 can be used to estimate the vibration
frequencies as well as harmonic coefficients.

40
Table 3.1: FCMW radar range and velocity estimations and their min/max detectable
bounds

Parameter Range Radial velocity Vibration frequency

fb c
 
b
Estimation R̂0 = 2K v̂ = λmax
2 argmaxf (|Ỹ (f )|)a fˆv = argmaxk p̂k 
cfb,max λmax 1
Max 4K 4Tc 2Tc
cfb,max λmax c
Min 2N K 2M Tc –
a
Ỹ (f ) is the FFT of ỹ[n] in 3.18.
b
p̂k is output power of the optimum filter for ωk .
c
the minimum can be anything.

In the following chapters, we will address the radar performance by designing different
tests. Then, the radar results for vital signs detection will be presented.

41
Chapter 4

System verification with experiments

and simulations

In this chapter, we will investigate the radar performance in terms of range, radial dis-
placement, and vibration detections. To that end, we designed three different tests as
follows:

ˆ Antenna chamber tests: The radar accuracy in the range, radial displacement,
and vibration detections are examined in a clutter-free environment i.e. an antenna
chamber.

ˆ Stationary tests: The phase received from a stationary target is investigated where
it has a linear trend. The source of the phase slope is identified and it has been shown
it does not deteriorate the vital signs detection.

ˆ Pendulum tests: The performance of the system is evaluated for vibration fre-
quency estimation by employing the algorithm 2. In these experiments, we designed
a pendulum system with which we will investigate the problem of DC compensation,
phase harmonics, and frequency estimator performance before discussing the vital
signs application in the next chapter. The pendulum system allows us to modulate
the phase synthetically similar to what is done by the cardio-respiration activities. In
addition, the subtleties during radar analysis, as mentioned in the previous chapter
theoretically, are investigated practically.

42
 (a) (b)

Figure 4.1: Antenna chamber

Table 4.1: Radar configuration for chamber tests

Parameter Tc Frame rate Chirp slope fb,max Na Mb

Values 57 µs 20 f ps 70 M Hz/µs 2 M Hz 128 256
a
Range FFT size
b
Second FFT size, which is either for velocity or vibration maps

4.1 Antenna chamber tests

The radar was placed in an antenna chamber which is, an ideal, a clutter-free environment
allowing to examine the reflections from the desired targets. The targets we chose are
two rectangular aluminium plates. One of them is 2 × 2 cm and the other is 20 × 30 cm.
The Radar configuration parameters are given in Table 4.1. The min/max values of each
variable corresponding to the values in Table 4.1 are given in Table 4.2. Since the minimum
resolvable distance difference is 3.76 cm, the small plate should only be seen in a single
range bin, in contrast, the big plate fills more than three range bins. In Figure 4.1, the
radar and the plate locations and their alignment are shown.
An FFT of size N is applied on each received chirp samples. Then the DC compensation
is applied. The range-velocity map is obtained by taking the second FFT over the DC
compensated range bins across M chirps. For range bin phase spectrogram map, instead
of taking FFF of the complex samples, the short-time Fourier transform (STFT) is taken
over unwrapped phase of the target range bin.
In figure 4.2a, the small plate is located at 1.46 m and occupying about 3 range bins.
Ideally, the target should be seen in a single range bin, as mentioned before, but due to

43
 Table 4.2: Min/Max value of different parameters

Parameter Range Velocity Vibration

Max 8.57 m 19.5 mm/s 10 Hz
Min (resolution) 3.76 cm 0.1522 mm/s 0.0781 Hza
a
FFT is used to calculate the vibration thus the minimum is the
FFT resolution.

phase noise of the FMCW generator, the beat signal does not have a sharp peak only at
fb (see 3.4). The spectrum widening is one effect of the phase noise reducing the range
accuracy. In fact, exp(jφn [m]) in 3.19 is multiplied to the phase noise-free part of ỹ[m]
in (3.18), which spreads the beat signal spectrum. Also note that there are reflections
at other ranges meaning that the chamber is not perfectly absorbing the wave at this
frequency. Figure 4.2b shows that everything, including the target, is perfectly still and
this is consistent with Figure 4.2c as well.
In comparison to Figure 4.2a, in Figure 4.3a the large plate range bin is wider as
expected and the plate is vibrating around its hanger in Figure 4.3b. This is not a radial
velocity as the range-velocity map does not have a single peak implying that there is
periodic movement. Figure 4.3c also indicates that the plate was moving with an almost
constant vibration frequency of 0.7813 Hz during the entire recording.
We intentionally tapped the plate to oscillate. Different maps for this case are illustrated
in figure 4.4. First of all, the range-velocity map is full of harmonics generated due to the
high magnitude of vibrations. Secondly, the phase spectrogram shows a smooth reduction
of the vibration harmonics over time. The fundamental frequency of the vibration is 0.7813
Hz. This is almost the same vibration fundamental frequency as in figure 4.3, but with
different magnitude. In addition, figure 4.5 depicts the time variation of the range bin
unwrapped phase together with the wrapped phase. Initially, the phase is not unwrapped
correctly because of sparse sampling of the phase. There is no upsampling in the processing
but it can be used to uncover the phase even for the beginning moments.

44
(a) Range-slow time map (b) Range-velocity map

(c) Range bin phase spectrogram

Figure 4.2: Different maps for the small still plate

45
(a) Range-slow time map (b) Range-velocity map

(c) Range bin phase spectrogram

Figure 4.3: Different maps for the large still plate

46
(a) Range-slow time map (b) Range-velocity map

(c) Range bin phase spectrogram

Figure 4.4: Different maps for the large high vibrating plate

47
 80

60 Wrapped
Unwrapped

40

20
Displacement (mm)

-20

-40

-60

-80

-100
0 20 40 60 80 100 120 140 160 180
Time (s)

Figure 4.5: Phase variation of the oscillating plate

48
4.2 Stationary tests
In a long-time run of the radar, the phase exhibits a linear trend, which could result in
a high DC in the phase. The trend does not depend on the range of the object and it is
indeterministic. The small slope in the phase means that there is a residual frequency in
the received complex signal that might be due to residual delay between Tx and Rx. The
residual delay might be either because of delays in the circuit path or non-linearities in the
FMCW signal generation. There is no nonlinearity in the FMCW output; therefore, the
reason for the linear phase trend is due to group delay variation of the PA and LNA over
temperature. Also, as the PA heats up, the group delay varies so that the phase slope is
varying from one experiment to another.
To show the repeatability of the hypothesis, the tests were run for different objects at
different ranges. The plates were free to have micro movements as they were hung up to
the chamber ceiling. The radar is configured with parameters in Table 4.1. For example,
figure 4.6 shows a target’s range phase, which was placed at 1.26 m. There is no need for
phase unwrapping as the phase changes in a few radians. The slope in this experiment is
about 2 µm/s. There are other peaks in the phase spectrum at 2.148 Hz and 8.438 Hz.
The phase magnitude for these peaks is within micro meter, for instance, a peak at -50 dB
is equivalent to 3.2 µm. It is important to notice that there is a noise floor for the phase,
which is -70 dB/Hz.
For vital signs detection, since the BR is as slow as 0.1 Hz, the phase DC should be
removed for BR estimation.

49
 -1

-1.2

Unwrapped
Displacement (mm)

-1.4

-1.6

-1.8

-2

-2.2

-2.4
0 50 100 150 200 250 300
Time (s)

(a) Phase time trace

-10

-20

-30
PSD (dB/Hz)

-40
X 8.438
-50 X 2.148 Y -53.26
Y -59.52
-60

-70

-80
0 1 2 3 4 5 6 7 8 9 10
Frequency (Hz)

(b) Phase spectrum

Figure 4.6: Linear phase trend

50
4.3 Pendulum tests
To understand the theoretical analysis done in the previous chapter, we designed a series
of experiments with a system containing two pendulums. The system has two pendulums
connected to two servo motors with which we can control the swing speed as well as the
swing angle. Figure 4.7 shows the base on which the servo motors and the wooden pen-
dulum arms are mounted. Two raspberrypi’s are used to generate pulse width modulation
(PWM) signal for servo motors to adjust the angle of the arms and their swing speed. The
end of each pendulum is covered with a metallic sheet to have higher reflections than the
other parts. The radar is placed at different distances to the base centre. In fact, the two
pendulums are at the same distance to the radar.
The servo motors could not move the arms smoothly. Particularly, the arms were
moved in a step-wise manner by the motors resulting in step-wise phase modulation. The
problem were more sever when the swing speed was very low or the swing angle was very
wide. Indeed, at the two ends of swing period, when the swing direction changes, the arms
were shaking resulted in a high order harmonics. The two issues where controlled by setting
appropriate values for the swing angle and the speed. Therefore, we could not expect a
fully-controlled smooth swinging, which is enough to model the chest wall movement.
The radar is tested by the pendulum system either when one or both pendulums were
moving. The later will be a reference for the vital signs detection in which we have two
independent vibrating objects at the same range.

Figure 4.7: Pendulum system

51
 0.04
Reconstructed
0.03 Original

0.02

0.01
Phase (Rad)

-0.01

-0.02

-0.03

-0.04
0 1 2 3 4 5 6 7 8
Time (s)

Figure 4.8: Phase time trace for the pendulum at 2.5 m.

4.3.1 One-pendulum tests

In this series of experiments, we will use a single pendulum to modulate the wave echo. We
placed the base at the range from 30 cm to 2.5 m when the pendulum was swinging with
almost 2.1 Hz. The radar configuration is the same as Table 4.1 with the valid values in
Table 4.2 except that the frame rate is 80 fps. Thus, the vibration rate can be at most
40 Hz. This is much higher than what we need, therefore, the signal is decimated before
passing it to the fundamental frequency estimator to have a sampling of 20 Hz.
The original waveform in figure 4.8 is extracted after phase unwrapping process (see
algorithm 1) and converting the real phase to the complex phase procedure described in
section 3.5.3. The pendulum was at 2.5 m away. The reconstructed waveform is ob-
tained by using the estimated complex coefficients in the model (3.36) with the triple
(N,M,L)=(1024,100,4) (in samples). The estimated waveform is very close to the original
proving the validity of the estimation even for the far vibrating objects.
As mentioned before, the pendulum was moving step-wisely with a jitter at end of each
swing period. As a result, the modulated phase is not purely sinusoidal, which is plotted in
Figure 4.8. Also, in Figure 4.9, the phase spectrogram is plotted for the two ranges. The
harmonics of the phase are pretty obvious for the pendulum at 30 cm, however, they are
less visible at 2.5 m. In fact, the phase noise decorrelation effect at higher ranges makes
the received phase noisier. As a result, the higher order harmonics buried under the phase
noise power in figure 4.9b.
After running the pendulum for a long time at each range, the accuracy of the fun-
damental frequency estimator in terms of number of probability of error versus range is

52
 (a) (b)

Figure 4.9: Phase spectrograms of a pendulum with vibration of 2.1 Hz at 30 cm and 2.5
m.

obtained in Figure 4.10 . The prominent fact is that as the range increases the phase
quality degrades due to the effect of phase noise uncorrelation (see (3.7)). Furthermore,
the thermal noise power is independent of the target range because the received power of
the far target is very small due high attenuation, the SNR is lower. Though, the phase
noise uncorrelation and SNR reduction deteriorate the phase quality.

DC offset cancellation for one pendulum

The DC in the received complex signal has three origins as mentioned before. In practice,
the DC is removed from the received beat signal due to antenna coupling and RF cross-
talk. Furthermore, the stationary clutter produces a DC which will appear in the recived
range bin. The pendulum system does not have large RCS to shift the complex signal
constellation from the origin1 . However, when a big object like a chair is placed next
to the pendulum, the constellation will shift according to the RCS of the chair. For the
illustration, the pendulum and the chair were placed at 1.5 m. The radar configuration
1
In fact, the non-vibrating parts of the pendulums’ base can be seen as a stationary clutter. More
specifically, the received range bin is a superposition of the two signals, one coming from the vibrating
target and the other one from the base body. Hence, if the stationary parts of the base have large RCS,
it means that that there will be a high DC in the received range bin.

53
 fv 2.1Hz with the tolerance of 3.738%
10-1
Error rate

10-2

10-3

1 1.5 2 2.5
Distance (m)

Figure 4.10: f0 estimation error rate for a single pendulum.

is the same as Table 4.1 except the frame rate is 200 fps. The pendulum frequency was
set to 6.5 Hz such that the phase was modulated with less harmonics. In Figure 4.11, the
received complex signal trajectory is depicted where the center of the phase modulation
is shifted back to the origin by estimating the center (and the radius) using (3.27). The
phase variation is shown in Figure 4.12 before and after DC compensation. Visually, the
DC removed signal has lower distortion.
The phase spectrograms in Figure 4.13 prove that the phase harmonics reduced. More
specifically, the power spectral density (PSD) of both signals are demonstrated in Fig-
ure 4.14. As the marker indicates, the second harmonic is reduced by 12 dB using DC
compensation.
This should be noted that the DC compensation is required for the phase analysis if
we know that there are stationary clutters at the range of interest, otherwise removing it
does not improve the phase quality and also it distorts the phase.

54
 3000
Before DC Correction
DC Corrected
2000
Estimated Center
Origin
1000

0
Imaginary

-1000

-2000

-3000

-4000

-5000
-4000 -3000 -2000 -1000 0 1000 2000 3000 4000
Real

Figure 4.11: Signal constellation with a shifted origin

4
Without DC compensation
3 With DC compensation

2
Phase (Rad)

-1

-2

-3

-4
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Time (s)

Figure 4.12: Phase modulation with a single pendulum before and after DC cancellation

55
 (a) With DC (b) Without DC

Figure 4.13: Phase spectrograms before and after DC offset cancellation

50
X 14.06
40 Y 34.09
Without DC compensation
30 With DC compensation

20
PSD (dB/Hz)

X 14.06
Y 21.92
10

-10

-20

-30
0 5 10 15 20 25 30 35 40 45 50
Frequency (Hz)

Figure 4.14: PSD of the phase signal before and after DC compensation

56
4.3.2 Two-pendulum tests
For the vital signs’ application, the chest wall moves by the respiration and heart beats.
To emulate the behaviour, we used the two-pendulum system as a reference for the radar
accuracy in detecting two independent vibrating targets at the same range2 . The magni-
tude respiration motion is about ten times greater than the heartbeats (see Ah , Ab in 3.22).
In addition, the frequency of BR is about 6 times less than HR. Considering these ratios,
the frequency for the two pendulums were set to 2.04 Hz and 6.5 Hz, respectively. Let us
call the pendulum with lower frequency p1, and the other p2. The amplitude of vibration
for p1 to p2 were set to 10.
There is no need for DC compensation. The radar parameters are the same as Table
4.1 except the frame rate is 200 Hz. The higher the frame is, the better the phase is
unwrapped. The received signal is down-sampled to the sampling of 50 Hz after phase
unwrapping. Furthermore, the fundamental frequency estimation parameters were set to
(N,M,L)=(10 seconds, 4 seconds, 4).
The pendulums were placed at the ranges from 85 cm to 2.5 m. The two figures in
4.15 show the single-side band phase spectrograms obtained at two different ranges. The
harmonics of p1 up to 4th order, about 8 Hz, can be seen in 4.15a. In contrast, the
harmonics of p2 decay faster such that they cannot be found higher than the second order.
Also, there are intermodulation components between p1 and p2 up to the third order. For
instance, there is a peak around 15 Hz, which is 2f1 + f0 where f1 and f0 correspond to
the frequency of p1 and p2, respectively. The noise level in the phase domain at 2.5 m for
the two pendulums is higher than the single pendulum (see Figures 4.9b and 4.15b). This
stems from the appearance of the intermodulation terms in the two-pendulum system.
The frequency estimator performance versus different ranges is depicted in Figure 4.16.
For the comparison, the error rate for the single-pendulum system is also included in the
figure. To obtain each point in the figure, it was required to run the system for more
than 40 minutes in order to measure the error rate above 0.02. The important fact is
that the error rate for the p2 is higher, which is justifiable for a phase modulation with
20 dB less power than p13 . Also note that the lower bound for the system performance
is the single-pendulum system. Specially, with this radar, the error rate of the vibration
2
From the biological point of view, the respiration and heart rates depend on each other such that an
increase in the BR will also increase HR. However, for the purpose of signal processing, we assume that
they are independent.
3
The power of the phase modulation for p2 is 20 dB less than p1 since its magnitude is 10 times less
than p1.

57
 (a) (b)

Figure 4.15: Phase spectrograms of the two pendulums at 85 cm and 2.5 m

frequency estimations in the two-pendulum system (as well as vital signs) is higher than
the single pendulum system by using the optimum filter.

58
 10 0

10 -1
Error rate

10 -2 f, single pendulum
f 1 , two pendulums
f 0 , two pendulums

10 -3

10 -4
0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
Distance (m)

Figure 4.16: Frequency estimator error rate for two-pendulum and single-pendulum exper-
iments at different ranges: f0 , f1 are for p1 and p2, respectively.

59
4.4 Matlab simulations
The Simulink environment of Matlab leverages the same toolbox functions and it provides
pipeline processing 4 . To be more clear, it does not run the blocks in different threads like
a multi-thread program, but it can use GPU or the CPU itself to parallelize the processing
whenever it is possible. The parallel processing basically depends on the intelligence of
Matlab to choose which parts can be run on different CPU cores separately. In general,
Simulink provides a higher simulation speed and has rich and powerful tools for visualizing
the signal properties, which help to understand, identify, and debug the program.
As a matter of fact, we used Simulink for analyzing the system under practical circum-
stances, which is based on the most realistic assumptions that has been used so far. We
used phased array and antenna toolboxes together to model an FMCW radar. The model
incorporates all the transceiver chain from the generation of the FMCW signal, transmit
hardware, Tx/Rx antennas, the channel, objects, down to the signal processing at the
receiver end. The first page of the designed Simulink model is illustrated in Figure 4.17.
The FMCW signal is generating by a ready block from the phased array toolbox. The
signal goes to a phased noise block, which adds a phase noise to the generator. Next, the
signal is fed into an antenna array, which propagates the equivalent fields in the space.
The field is attenuated by considering the free space channel model and it is reflected by
three scatterers. The scatterers are placed at the same distance on purpose. Two of them
model the vibrations of the two vibrating objects while the third one models the stationary
clutter. After the field bounces off the objects, it is collected by the receiver array, which
is at the same location as the transmitter. In the wave propagation, the wave polarization
is considered in order to count the effect of the polarization change. Then, the thermal
noise will be added and the mixture is correlated with the transmit signal to get the beat
signal. The beat signal has a very high sampling rate but it has a low IF frequency. Hence,
it is downconverted to 2 MHz sampling rate5 . We denoted this process as the ADC block
in Figure 4.17. After ADC, the signal is delivered to the signal processing block, which
performs chirp collection, phase unwrapping, and fundamental frequency estimation.
Figure 4.18 shows the schematic of the radar and targets’ platform positions. The
propagation model is based on the ray tracing model. The model gets the propagation
direction from the transmitter to the targets and it applies the antenna gain and phase on
the signal. The targets can have a scattering matrix, which can be a time varying property.
4
different from parallel processing
5
Note that the same baseband sampling rate as fb,max in table 4.1. Also, the downsampling factor is
50, which has to be carried out by multi-chain downconverter to avoid of having long filters. Actually this
is achieved by using 3 stages of downsampling by the factors of 5,5,2, respectively.

60
 Open Configuration
parameters ... mm-wave FMCW Radar with two vibrating objects
Antennas
FMCW Generator & Targets & Channel
Hardware
propagation
[FMCW_gen] Channel
[6400x1]

double [3x1]
double (c) [6400x1] Phase double (c) [6400x1] double (c) [6400x3] Pos TargetPos
double (c) [6400x3]
FMCW Target1 [3x1]
[6400x1] Noise [6400x1] double [3x1]
[6400x3]
Vel TargetVel
[3x1]
double (c) [6400x3]
Sig SigAtRx
[6400x3]
double (c) [6400x3] double [3x1]
Target2 Pos2 TargetPos2
[6400x3] [6400x3] [3x1]
double [3x1]
double [3x1] [3x1] double [3x3] Vel2 TargetVel2
RadarPos double (c) [6400x3] double (c) [6400x1]
Pos SigAtRx [SigAtRF] [3x1]
[3x1] [3x3] double [3x1]
possible of adding an [6400x3] [6400x1]
Platform 2 Pos3 TargetPos3
double [3x1] [3x1] double (c) [6400x3]
extra delay to acount for [3x1]
Vel Target3 double [3x1]
double [3x3]
the RF delay [6400x3] Vel3 TargetVel3
[3x1]
RadarVel [3x1]
[3x1] [3x3]
2
[3x1]

Scaling up the signal to

avoid of having a very
small signal [6400x1] [6400x1]
RefX AWGN [SigAtRF]
double (c) [6400x1] double (c) [6400x1]
[128x1] [6400x1] [6400x1] [6400x1]

61
I/Q
Signal Processing double (c) [128x1]
filtering and decimation double (c) [6400x1] double (c) [6400x1] double (c) [6400x1]
Dechirp
Imbalance [6400x1]
X [FMCW_gen]
double (c) [6400x1]

BaseBand ADC to 2 Msps Receiver

Figure 4.17: Matlab Simulink main page

 Figure 4.18: The radar and the targets’ positions in the simulation

Without loss of generality, we assumed that they are constant. Although they are fixed,
the mean RCS value of each target is different. The first target, denoted as T1 in the
figure 4.18, is aimed for modelling the lung cavity for the respiration. Particularly, the
second target (T2) has separate RCS, vibration magnitude and frequency to demonstrate
the heart activity. The last target (T3) will help to give us insight about the effect of a
stationary clutter on the phase quality.
We chose a 4 by 2 Tx/Rx arrays in the simulations. They are uniform linear array
(ULA)s with dipole antenna elements and with half-wavelength spacing. Their patterns
are shown in the schematic of 4.18. They were designed by the ready models in the antenna
toolbox of Matlab. There is no restriction of using a custom antenna module by passing
required parameters to Matlab antenna system object. The object can be replaced with
the one that is used here and it will carry out all the necessary processing for converting
a signal to a propagating field.
The actual radar system has a sweeping bandwidth of 4 GHz at 77 GHz. The simulation
for that sweeping bandwidth requires to generate a signal with a sampling rate of 4 Gsps.
In fact, the number of samples per 20µs chirp will be 80,000 samples. As a result, the
size of the vectors and the processing time will be increased to an amount with which the
simulation will not be made in a reasonable order of time. Thus, for the purpose of the
simulation and without loss of generality, the sweeping bandwidth was set to 100 MHz
by the chirp slope of 1.5625 M Hz/µs and the duration of 64 µs. The order of the valid
range, velocity, and the vibration frequencies changed to values in Table 4.4. Also Table 4.3
summarizes the simulation configuration’s parameters. The magnitude of the first and the
second vibrating object are set with a ratio of 10 in order to mimic the situation between

62
the breathing and the heart waveform. Also, their frequencies have a ratio of 6 like the ratio
between the HR to BR. The first target, v1, has the frequency of 24 Hz while the second
one has the frequency of 105 Hz. In the entire simulation, the DC compensation is ignored
due to absence of any stationary clutter. The phase noise is modelled by using a ready
Matlab block from the communications toolbox. We set the noise level for the frequency
shifts of [500 KHz, 1 M Hz] to [x, −91] dBc. The first phase noise level is varying in
different simulations and we mentioned that on each simulation (see Figure 4.19). The
second phase noise level is equal to the TI radar phase noise reported in Table 2.1. 500
KHz frequency shift from the carrier means that the target is at 47.9667 m away from
the radar. Accordingly, 1 MHz shift means that the target is at 95.9336 m. In mm-wave
frequencies the wave is attenuated by a factor of 10−13 , but for the simulation purpose,
we added the thermal noise by a controlled SNR to be able to test for different SNRs.
Otherwise, in practice, the receiver could not deliver the signal with a high SNR of 10 dB
at these ranges.
The SNR is set such that the desired range experiences that SNR value. Therefore, the
noise power level is computed by having the power of the echo signal and considering the
target distance, antenna gains, and the free-space loss as well as the number of range FFT
bins. The later is required to have the effective SNR in the range bin as they are reported
in the results. The number of range FFT bins is 128 (see Table 4.3), so the noise power is
multiplied to 128. This factor causes the effective desired SNR in the range bin bandwidth,
which is a 128th portion of the baseband bandwidth. In practice, each range bin can have
10log10 128 = 21.072 dB higher SNR with respect to the total baseband bandwidth. For
instance, if we set SNR = 10 dB, then the actual SNR for the received beat signal can be
-11.0720 dB.
The spectrum of the phase without applying DC compensation for two SNRs are shown
in Figures 4.20 and 4.21. A comparison of the two figures reveals that the noise floor of the
phase spectrum will be increased by reducing the SNR. This fact is analytically expressed
in the equation of (3.35). The total harmonic distortion (THD) analysis of up to 6 order
harmonics of the first vibrating object shows almost 5 dB THD increment in the phase
domain, which is in agreement of the phase error variance in (3.35). The constellation
points of the complex signal are shown in Figure 4.22. The uncertainty in the trajectory
is increased by reducing the SNR. In addition, there is a DC offset around which the
constellation points cloud is formed. This DC value is due to the phase modulation of the
vibrating objects [43] and it does not distort the phase. The time domain of the phase
for two values of the phase noise level 1 are plotted in figure 4.23. The phase variance is
increased as a consequence of the SNR increment.
Indeed, the simulink model was run for a long time to obtain the frequency estimator

63
 Table 4.3: Radar configuration for the Simulink model

Parameters Value Description

Ts 64 µs
Tr = 4Ts 256 µs Collecting the chirps every 4 chirps
Chirp slope 1.5625 M Hz/µs
Sweeping BW 100 MHz
fb,max 2 MHz
λmax 3.9 mm
N 128 Range FFT size
NDop 128 Vibration FFT size which is equivalent to 32.76 ms.
R0 , R1 , R2 14.9896 m Equivalent to the range bin index of 11.
Av1 , Av2 0.4, 0.04 mm Magnitudes of the first and the second vibrating object
fv1 , fv2 24, 105 Hz Keeping the ratio of 6 between the two vibrating objects’
frequency
L 4 System model order for f0 estimation
M 81 The f0 estimation parameter
Nf 0 203 The f0 estimation parameter

Table 4.4: Radar parameter detection valid ranges for the Simulink model

parameters Range (m) Velocity Vibration frequency

Max 95.93 3.8 (m/s) 1.953 KHz
Min 1.499 5.94 (cm/s) 30.51 Hz

64
 -40
n,1
=-50 dBc

-50 n,1
=-80 dBc
Magnitude (dB)

-60

-70

-80

-90

-100
10 -2 10 -1 10 0 10 1 10 2
Frequency (MHz)

Figure 4.19: Phase noise block frequency response for the minimum and maximum phase
noise level 1. It is important to note that the x-axis is plotted in logarithmic scale in order
to be able to depict the magnitude response transition. Otherwise, in the linear scale, the
magnitude drops very rapidly like a Dirac function.

performance for different SNRs and different phase noise levels (figure 4.24).
Firstly, in Figure 4.24a, the gap between the estimation error of f1 and f0 is obvious.
Secondly, if the desired Perr is 10−26 , then the SNR should be greater than or equal to 16
dB such that the condition is satisfied for both f0 , f1 . This means that if the antenna gains
are 10 dB at both Tx and Rx, and the receiver NF is 16 dB7 , the maximum detectable
distance will be 1.667 m by using equation (3.9). Thirdly, there is an error floor for high
enough SNR. The floor is due to the phase noise. Therefore, in order to have the error rate
less than 10−2 , the phase noise level should be less than or equal to -80 dBc. Note that
the curves in Figure 4.24a are not stretched beyond of 10−2 due to having small number
of estimation samples8 . Thus, the error floor for f0 is not shown since it is less than 10−2 .
The estimation accuracies versus the phase noise levels are plotted in the figure 4.24b.
The phase noise level has a greater influence on the performance of f0 than f1 (compare
the no-noise cases). For all SNRs, the error rate at the phase noise of -50 dBc is almost 1.
In addition, the figure confirms that for having an error rate of less than 10−2 , the phase
noise level at 500 KHz offset should be less than -80 dBc.

6
meaning that out of 100 trials, one of the estimations are wrong
7
this is the NF of the AWR1642/AWR1443.
8
Each pair of graph in Figure 4.24 for f0 , f1 requires 8 hours of simulation run.

65
Figure 4.20: Phase spectrum for SNR = 10 dB and the phase noise level 1 = -80 dBc

Figure 4.21: Phase spectrum for SNR = 5 dB and the phase noise level 1 = -80 dBc

66
 10
0 0
0
-10 -10
Quadrature Amplitude

Quadrature Amplitude

Quadrature Amplitude
-10

-20 -20 -20

-30 -30 -30

-40 -40
-40
-50
-50 -50
-60
-60
-60
-70

-60 -50 -40 -30 -20 -10 0 -60 -50 -40 -30 -20 -10 0 -60 -40 -20 0
In-phase Amplitude In-phase Amplitude In-phase Amplitude

(a) SNR = 18 dB (b) SNR = 10 dB (c) SNR = 5 dB

Figure 4.22: The constellation of the complex signal at different SNRs and the phase noise
level 1 of -80 dBc: the color shading is getting stronger for the most recent samples and
the origin of the plane is denoted by the square.

1.5
n,1
= -50 dB
SNR = 18 dB = -80 dB
n,1
1
Phase (Rad)

0.5

-0.5

-1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
Time (s)

Figure 4.23: Phase time trace for two different phase noise levels

67
 Frequency estimator performance
10 0
f0, no phase noise
f1, no phase noise
f0, n,1 = -80 dB
f1
f0, n,1
= -70 dB
f1, n,1
=-80 dB
f1, n,1
= -70 dB
X 12
Y 0.125
f0
Perror

10 -1
X 15
Y 0.09

10 -2
0 2 4 6 8 10 12 14 16 18
SNR (dB)

(a)

Frequency estimator performance

10 0

10 -1
f0, SNR = 0 dB
f1, SNR = 0 dB
f0, SNR = 18 dB
Perror

f1, SNR = 18 dB
f0, no noise
f1, no noise
10 -2

X -70
Y 0.005

10 -3
-80 -75 -70 -65 -60 -55 -50
Phase noise level 1 (dB)

(b)

Figure 4.24: Fundamental frequency estimation error rates vs SNR and phase noise levels:
φn,1 is the phase noise level of 1. 68
Chapter 5

Application on vital signs detection

Although in general one can use the theories developed in Chapter 2 to detect and find
one or two vibrating objects and their features, specifically it can be used for vital signs
detection. In this chapter, we will investigate the mm-wave FMCW radar application for
vital signs detection. For this purpose, we will use the findings in Chapter 4.

5.1 Signal processing flow

In Figure 5.1, the block diagram of the proposed signal processing chain for vital signs
detection is shown. After sampling the beat signal with fb,max , the range FFT is applied
over the samples of each chirp and the result is a vector, which is called complex range
profile. By collecting consecutive complex range profiles from multiple chirps and putting
them into a matrix in a row-wise manner, the range-slow time matrix is constructed with
M rows (i.e. M chirps).
Prior to taking the angle of the received complex signal, we must make sure that any
non-linearities, distortions, and artifacts have been removed since the phase computation
is highly non-linear and it increases the complexity of removing those imperfections. For
instance, the DC value of both in-phase and quadrature components of the complex signal
must be removed otherwise, it will affect the phase quality. There is no high DC component
in the received complex beat signal since none of the DC generators are present at a “good”
range except the one that is generated by the cardio-respiratory phase modulation itself. A
“good” range is the first range at which we have an intersection to the body. For example, if
the subject side is faced to the radar, then his/her shoulder has an intersection to a sphere

69
 Figure 5.1: Proposed signal processing chain

surface, which has the smallest radius (radius here means the range). In this way, we
minimize the effect of the stationary parts of the body, which are recognized as stationary
clutters with high RCS. In other words, we do not need to use DC compensation in the
following scenarios for the vital signs tests.
After selecting the desired range bin as described before, we upsample the complex
signal in order to add redundant samples to have a higher chance of correct phase unwrap-
ping. The upsampling factor for each test will be given. The upsampled complex signal is
ready to go to the phase domain.
The phase of each column in the range-slow time matrix is calculated by using tan−1 (.)
so that the output phases are wrapped in [−π, π]. In contrast, the phase can change beyond
of ±π because x(t), the physical displacement, can be greater than λmax /4. Therefore,
there should be a mechanism to unwrap the phase beyond of ±π as it was mentioned
in section 3.4. Phase unwrapping is performed on the columns of the range-slow time
matrix separately (Figure 5.1). For a detailed analysis on the phase unwrapping process
see algorithm 1.
The unwrapped phase is downsampled to 40 Hz. Then the signal goes through real to
complex transformation in which its sampling rate is reduced by a factor of two making the
final sampling rate of 20 Hz. The final sampling rate is enough for the vital signs detection
since it contains up to the fourth order harmonics of the HR (it is assumed that the
maximum HR is 2 Hz at rest condition). The DC value of each column is removed during
real to complex transformation. Then, to find vibration frequencies, the fundamental
frequency estimator is employed separately for each vital sign (see algorithm 2).

70
 Radar configuration is the same as in Table 4.1 except that the frame rate may vary
from one experiment to another. Thus, for the valid values of the range, Table 4.2 is a true
reference. The valid range of vibration frequencies depend on the frame rate and it can be
computed using Table 3.1.
In the following sections we will discuss the results and the comparison of this work to
similar works using FMCW radar.

5.2 Experiments’ setup

In Figures 5.2 and 5.3 the positions of the subject and the radar for sitting and sleeping
tests are shown. In the sleep experiments, the radar platform was mounted on the ceiling
above the bed as shown in Figure 5.3. We will demonstrate the results for both set ups. The
radar was connected to a PC via a USB cable to control the reception and transmission.
The raw data is taken and analyzed in Matlab with the processing chain shown in Figure
5.1.
In all tests, a male subject sat or slept approximately 1.5 to 2 m in front of the radar
on a chair or on a bed. The 28-years-old subject has the height of 1.7 m and the weight of
62 kilograms. In the sleeping tests, the subject was free to sleep in different positions such
as back, stomach, right, and left sides. The Hexoskin vest is worn by the subject to track
his BR, HR, and their corresponding waveforms.

71
72
Figure 5.2: Sitting tests setup
73
Figure 5.3: Sleep tests setup
5.3 Waveforms
The vital signs waveform conveys information about the subjects health condition. In
fact, the respiration and ECG signals help to diagnose potential diseases in the subject.
For instance, there are enormous studies on the heart waveform, ECG, to understand
the emotional state of an individual. This analysis is not possible except by having the
required cardio-respiration waveforms. Here, we demonstrate the potential use of FMCW
radars for deep clinical health diagnostics. The frame rate of the radar was set on 80 fps
and we did not use upsampling. For example, Figure 5.4 shows the waveform in which
the subject was sitting on a chair faced to the radar1 . Initially, he was told to hold his
breath to capture the phase modulation only due to the heartbeats. After some time, he
released his breath to have the combination of the breathing and heart waveforms. The
wrapped phase is also drawn to show the effective phase unwrapping. The magnitude of
the phase waveform for the breathing is almost 10 times less than the breathing, which
is equivalent to 20 dB less power. For a typical adult, the chest moves about 1-12 mm
and 0.01-0.5 mm due to breathing and heartbeat, respectively [29]. In fact, in practice,
the heart rate extraction is a great challenge due to having imperfect DC cancellation and
the strong respiration. Each peak in the heartbeat-only period corresponds to the R peak
in the ECG waveform, which is a sharp peak. The detection of the peak further helps to
identify the HRV parameter. From the respiration waveform2 , one can retrieve the minute
ventilation as well as the tidal volume. Although, they can be computed by the waveform,
our focus in this work is on the vital signs rates. Also consider that, there is a linear very
low slope in the phase generated by the time delay variations of the RF components, which
were warming up at the beginning of the radar operation. In fact, this behaviour will be
diminished after 5 minutes or so when the RF part goes in the stable temperature period.
In the sleep study, the radar frame rate was set to 200 fps and the upsampling factor was
2. The higher frame rate is used to be able to track possible body motions. In Figures 5.5
and 5.6, the vital signs waveforms are plotted against that of the Hexoskin vest for different
sleep positions. The waveforms all have almost peak-to-peak value of 3 mm except the
back-side waveform. This means that the body motion due to the expiration, inspiration
and the heartbeats is less from the back side. In these figures, the breathing waveform
reconstructed from the optimum filter design is plotted (see section 3.5). The reconstructed
noise-free waveform is the summation of the harmonics after estimation of a coefficients in
equation (3.36). The optimum filter design parameters for the breathing waveform were set
to (N,M,L)=(25 seconds, 10 seconds, 1). Actually, only the first order approximation of the
1
The sitting-on-the-chair tests were done in which the target is faced directly to the radar front side.
2
After releasing the breath, the dominant waveform is the breathing.

74
 25
Received
20
Unwrapped
15
Phase (Rad)

10

-5

-10
0 10 20 30 40 50 60 70 80 90
Time (s)

6
Received
4
Unwrapped

2
Phase (Rad)

-2

-4

-6
0 5 10 15 20 25
Time (s)

Figure 5.4: Vital signs waveform for hold-and-release-breath tests

75
breathing waveform is used since the waveform shape is very close to the pure monotone
signal. In contrast, for the heart waveform, the filter parameters were set to (N,M,L)=(25
seconds, 10 seconds, 8) because, ECG signal has high order harmonics, hence, we need a
high order system model. The results for the heartbeat waveforms obtained by the radar
and Hexoskin are shown in figures 5.7 and 5.8. Depending on the quality of the received
signal, at particular moments, the heart signal has closer shape to ECG.

76
 2 Radar
Back side
Hexoskin
1.5
Reconstructed
1
Displacement (mm)

0.5

-0.5

-1

-1.5

-2
0 10 20 30 40 50 60
Time (s)

(a) Back

2 Radar
Right side
Hexoskin
1.5
Reconstructed
1
Displacement (mm)

0.5

-0.5

-1

-1.5

-2
0 10 20 30 40 50 60
Time (s)

(b) Right

Figure 5.5: Vital signs waveforms for the sleep tests on the back and right side: in each
figure, the original phase waveform together with the Hexoskin breathing waveforms and
the breathing waveform reconstructed from the optimum filter are plotted.

77
 6 Radar
Belly Hexoskin
5 Reconstructed
Displacement (mm)

-1

0 10 20 30 40 50
Time (s)

(a)

2
Radar
Left side
1.5 Hexoskin
Reconstructed
1
Displacement (mm)

0.5

-0.5

-1

-1.5

-2
0 10 20 30 40 50 60
Time (s)

(b)

Figure 5.6: Vital signs waveforms for the sleep tests on the belly and left side: in each
figure, the original phase waveform together with the Hexoskin breathing waveforms and
the breathing waveform reconstructed from the optimum filter are plotted.

78
 0.2 Reconstructed
Back side
Hexoskin
0.15

0.1
Displacement (mm)

0.05

-0.05

-0.1

-0.15

-0.2
0 2 4 6 8 10 12 14 16 18 20
Time (s)

(a) Back

0.15 Reconstructed
Right side Hexoskin

0.1
Displacement (mm)

0.05

-0.05

-0.1

-0.15
0 2 4 6 8 10 12 14 16 18 20
Time (s)

(b) Right

Figure 5.7: Heart waveforms for the sleep tests on the back and right side: in each figure, the
ECG signal from Hexoskin vest and the heart waveform reconstructed from the optimum
filter are plotted.

79
 0.4
Belly Reconstructed
Hexoskin
0.3
Displacement (mm)

0.2

0.1

-0.1
0 2 4 6 8 10 12 14 16 18 20
Time (s)

(a) Belly

Reconstructed
0.15 Hexoskin
Left side

0.1
Displacement (mm)

0.05

-0.05

-0.1
0 2 4 6 8 10 12 14 16 18 20
Time (s)

(b) Left

Figure 5.8: Heart waveforms for the sleep tests on the belly and left side: in each figure, the
ECG signal from Hexoskin vest and the heart waveform reconstructed from the optimum
filter are plotted. Note that the ECG signal when the subject was sleeping on his belly is
very noisy, which is due to improper electrodes contact of the Hexoskin vest to the skin
surface. The R peaks are obvious to compare with the radar.

80
5.4 Respiration and heart rates

Breathing comparison
20
Breathing (per min)

15

B
e
= 0.65637/min
10
CV = 3.9892% Radar
Ricc=0.93592 Hexoskin

5
0 500 1000 1500 2000 2500
Time (s)

(a) BR time trace

Heart rate comparison
120
H
= 3.0193/min Radar
110 e
Hexoskin
Heart rate (per min)

CV = 3.6022%
100 Ricc=0.79757
90

80

70

60
0 500 1000 Time (s) 1500 2000 2500

(b) HR time trace

22 120
P = 2250 P = 2250
20
r = 0.94 r = 0.80
110
18 y = 1.00 y = 0.98

16 100
Radar HR
Radar BR

14
90
samples
12
linear fit
10 "y=x" line 80
samples
linear fit
8
70 "y=x" line
6
6 8 10 12 14 16 18 20 22 70 80 90 100 110 120
Hexoskin BR Hexoskin HR

(c) Radar BR vs Hexoskin BR (left), radar HR vs Hexoskin HR (right)

Figure 5.9: Radar BR comparison with Hexoskin (a), and radar HR comparison with
Hexoskin (b)

81
 In a 40-minute trial the BR and HR are estimated based on algorithm 2. During the
entire test, the subject was sleeping on his back meaning that his face was toward the
radar. After removing out-of-norm estimates, which can be due to spark noises and any
other external causes, the time trace along with the statistics of BR and HR are shown in
Figure 5.9a and Figure 5.9b. In these figures, the instantaneous values of the radar and
Hexoskin estimates are compared.
To have intuitions, some statistical parameters are defined to evaluate the similarity,
robustness and confidentiality of the radar estimations with respect to Hexoskin estimates.
If e is an error vector with the elements of ei , then the standard deviation of the error can
be stated as: v
u
u1 X P
σe = t (ei − ē)2 =k e k, (5.1)
P i=1
¯ and k . k are the notations for statistical mean and
which is the standard deviation where (.)
second norm of a vector respectively and P is the number of data points. The correlation
of variation (CV) is another factor represents the ratio between the standard deviation and
the mean value of the data, d:
σe
CV = ¯ × 100 (5.2)
d
One useful parameter is Pearson correlation coefficient denoted as Ricc :
PP ¯r dh − d¯h



dr − d
Ricc = qP i=1 i i i i

, (5.3)
P
dr − d¯2 dh − d¯2

PP
i=1 i ri i=1 i hi

in which dri and dhi are the i’th radar and Hexoskin estimations respectively. The value of
Ricc is in the interval of [−1, 1]. The value of 1 means the two data sets are fully correlated
and 0, in contrast, means they are uncorrelated. In fact, this correlation describes how
two data sets are similar. All the metrics are annotated in the time trace Figures of figure
5.9a and Figure 5.9b.
The sample points cloud in radar-Hexoskin planes is plotted in Figure 5.9c (left) and
Figure 5.9c (right) in which r is Ricc . The ideal case is when all the samples are sitting
on y = x line indicating that the radar data are the same as Hexoskin. In practice, the
BR data follows a similar linear behavior around y = x line with a slop of 1 and a 0
bias. Similarly, for HR, the linear fitted curve has a slop of 0.98 and a bias of 0. In the
acceptable ranges of HR values, the linear fitted curve is very close to y = x line while
they can diverge for very lower or very higher invalid values. It is also good to mention

82
Table 5.1: HR and BR error rates for different sleep positions: the underlined values are
the worst performances

Side HR (%) BR (%)

Left 5 4.26
Right 5.86 4.14
Back 3.95 3.39
Belly 2.5 1.54

Table 5.2: Estimation accuracy comparison to similar works

Ref. Distance (m) f min (GHz) HR (%) BR (%) pout (mW)

[15] 2.48 9.6 55.2 — 25
[66] 1 80 87.2 91.08 0.5
Ours 1.7 77 94 95 15.8

that the BR and HR values are well distributed in the whole range of vital signs to show
the agreement between the radar observation and the reference sensor.
In table 5.1, the error rates for the vital rates are listed for different sleep positions.
Each BR value of the radar, which was in the ± 1 bpm, was considered as a correct
estimation. For HR, the acceptable tolerance was set on ± 3 bpm. The table shows that
with the radar the accuracy of 95% and 94% for BR and HR is achieved, respectively.
The performance of our system is compared to similar works in Table 5.2. The values for
the front side tests in [66] are in the table to make a fair comparison to our experiments. The
performance reported in [66] has a high variance for a particular radar-target orientation.
However, we consider the average performance while there might be a case the values are
less than 4%. The system performance in [15] is not better than [66] and the best correct
estimation rate they obtained is considered here. BR accuracy for [15] was not reported
but our system shows an improvement in comparison to [66]. Our HR estimation is much
more accurate than [15] and it is so close to [66]. Besides, the output power of our radar is
higher than the others (last column of Table 5.2) but the target is somehow in a close range.
Although we could not change the distance due to the space limitation in the bedroom, it
is possible to use the radar for further distances.

83
Chapter 6

Summary, outlook, and future

extensions

In Chapter 3, the fundamental FMCW radar principles were discussed. In that, the DC
compensation method is investigated. But later, in Chapter 4, we have shown that there no
DC compensation is required if there is no big stationary target in the range. Also we used
the mathematical derivations in Chapter 3 throughout the experiments and simulations in
Chapters 4, 5. The experiments showed a good match between theory and the practice.
Indeed, the phase unwrapping process is proposed followed by the harmonic analysis based
on MUSIC. In fact, MUSIC was used to extract and enhance the periodic signals. The
enhancement of the heart signal was shown in the Chapter 5.
Chapter 4 verified the proposed system with the design of different experiments. They
are stationary tests, pendulum tests, and the Matlab simulation as well. We discussed the
results obtained by the simulations, which are very close to the practical observations.
In Chapter 5, the experiments results were provided for the vital signs detection. Gen-
erally, two sets of experiments were conducted. In one setting, a subject was sitting in front
of the radar facing to the radar at a right angle. In the second set of experiments, a subject
was sleeping on a bed and the radar was mounted on the ceiling above the bed. The final
results of BR and HR accuracies are compared to the preceding works. Our measurements
showed that HR and BR accuracies in the worst case for all sleeping positions were 94%
and 95%, respectively.

84
6.1 Future extensions
One benefit of the FMCW radars is their ability in range discriminations or localization.
Thus, it is straightforward to use them for multiple subjects monitoring. Suppose that a
single radar is attached on the corner of a bedroom and it monitors the heart and breathing
rates of the residents. This is feasible because it has been shown that the radar could detect
the vital signs from different positions to the targets.
TI radars can be used to make a network of the sensors. The radars can be cascaded in
the master-slave configuration suggesting that multiple radars can be used to continuously
monitor the health condition of people in an environment such as in an air plane.
We measured the BR and HR, but it is possible to measure the minute ventilation,
HRV etc.
For body motion cancellation, the use of a secondary radar has been suggested by
[43, 70]. This idea can be extended to FMCW radars as well. By placing another radar
under the bed, it is possible to eliminate the subject movements by giving the second radar
signal as a reference to an ANC unit.
In addition, a phased array antenna can be designed to narrow the antennas pattern.
This will help to have a focused beam on the body where the chest is moving rather than
having a wave impinged on the whole body. The other parts of the body, which are not
contributing to the vital phase modulation will be observed as stationary clutters. Thus,
exploiting narrowband arrays at both Tx and Rx help to resolve the DC issue in the
phase analysis. More advancely, TI has another type of mm-wave chips with the name of
AWR1242 [1] with which we can connect a couple of them to have a large of number of
antennas. In fact, the chipsets are only RF front-ends and we can make a phased array.
Using these chips can be a quick solution for developing a large array of antennas for a
mm-wave FMCW radar.
In general, based on the analysis carried out in Chapter 3, and simulations and experi-
ments in Chapter 4, we can design the FMCW radar hardware based on the requirements
obtained for the vital signs application. For instance, the phase noise should be smaller
than -80 dBc at the particular distance of the subject. The detectable range can be in-
creased by reducing the NF of the receiver and increasing the output power.

85
References

[1] AWR1243 76-to-81ghz high-performance automotive MMIC | TI.com.

[2] AWR1443 Single-Chip 76-to-81ghz Automotive Radar Sensor Integrating MCU and
Hardware Accelerator | TI.com.
[3] AWR1443boost AWR1443 Single-Chip 76-to-81ghz Automotive Radar Sensor Evalu-
ation Module | TI.com.
[4] AWR1642 Single-chip 76-to-81ghz automotive radar sensor integrating DSP and MCU
| TI.com.
[5] AWR1642boost AWR1642 single-chip 76-to-81ghz automotive radar sensor integrating
DSP and MCU evaluation module | TI.com.
[6] The BioRadio Wireless Physiology Monitor |.
[7] Computing the discrete-time ”analytic” signal via FFT - IEEE Journals & Magazine.
[8] The dielectric properties of biological tissues: III. Parametric models for the dielectric
spectrum of tissues - IOPscience.
[9] An electromagnetic model of human vital signs detection and its experimental valida-
tion - IEEE journals & magazine.
[10] Joint High-Resolution Fundamental Frequency and Order Estimation - IEEE Journals
& Magazine.
[11] Optimal carrier frequency of non-contact vital sign detectors - IEEE conference pub-
lication.
[12] A robust and computationally efficient subspace-based fundamental frequency esti-
mator - IEEE journals & magazine.

86
[13] Robust doppler radar demodulation via compressed sensing - IET journals & maga-
zine.

[14] Alfred V. Aho and Jeffrey D. Ullman. Foundations of Computer Science. W H Freeman
& Co, new edition edition edition.

[15] L. Anitori, A. de Jong, and F. Nennie. Fmcw radar for life-sign detection. In 2009
IEEE Radar Conference, pages 1–6, May 2009.

[16] S. Ayhan, S. Scherr, A. Bhutani, B. Fischbach, M. Pauli, and T. Zwick. Impact of

Frequency Ramp Nonlinearity, Phase Noise, and SNR on FMCW Radar Accuracy.
IEEE Transactions on Microwave Theory and Techniques, 64(10):3290–3301, October
2016.

[17] Constantine A. Balanis. Modern Antenna Handbook. John Wiley & Sons, Incorpo-
rated. Google-Books-ID: Q9OgkQEACAAJ.

[18] S. Benchikh, H. Arab, and S. O. Tatu. A novel millimeter wave radar sensor for med-
ical signal detection. In 2018 IEEE International Microwave Biomedical Conference
(IMBioC), pages 142–144.

[19] M. C. Budge and M. P. Burt. Range correlation effects in radars. In The Record of
the 1993 IEEE National Radar Conference, pages 212–216, 1993.

[20] A. T. Cemgil. Bayesian Music Transcription. [S.l. : s.n.].

[21] K. W. Chan and H. C. So. Accurate frequency estimation for real harmonic sinusoids.
11(7):609–612.

[22] N. Chernov and C. Lesort. Least squares fitting of circles. 23(3):239–252.

[23] M. G. Christensen and A. Jakobsson. Optimal filter designs for separating and en-
hancing periodic signals. 58(12):5969–5983.

[24] M. G. Christensen and A. Jakobsson. Optimal Filter Designs for Separating and
Enhancing Periodic Signals. IEEE Transactions on Signal Processing, 58(12):5969–
5983, December 2010.

[25] M. G. Christensen, A. Jakobsson, and S. H. Jensen. Fundamental Frequency Estima-

tion using the Shift-Invariance Property. In 2007 Conference Record of the Forty-First
Asilomar Conference on Signals, Systems and Computers, pages 631–635, November
2007.

87
[26] Mads Christensen and Andreas Jakobsson. Multi-Pitch Estimation. Morgan and
Claypool Publishers.

[27] Mads Christensen and Andreas Jakobsson. Multi-Pitch Estimation. Morgan & Clay-
pool Publishers, March 2009. Google-Books-ID: DtJcAQAAQBAJ.

[28] L. Ding, M. Ali, S. Patole, and A. Dabak. Vibration parameter estimation using
FMCW radar. In 2016 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), pages 2224–2228, March 2016.

[29] Amy Diane Droitcour. Non-Contact Measurement of Heart and Respiration Rates
with a Single-Chip Microwave Doppler Radar. 2006.

[30] N. Du, K. Liu, L. Ge, and J. Zhang. ApneaRadar: A 24ghz Radar-Based Contactless
Sleep Apnea Detection System. In 2017 2nd International Conference on Frontiers of
Sensors Technologies (ICFST), pages 372–376, April 2017.

[31] C. Gabriel, S. Gabriel, and E. Corthout. The dielectric properties of biological tissues:
I. literature survey. 41(11):2231.

[32] C. Gabriel, S. Gabriel, and E. Corthout. The dielectric properties of biological tissues:
I. Literature survey. Physics in Medicine & Biology, 41(11):2231, 1996.

[33] S. Gabriel, R. W. Lau, and C. Gabriel. The dielectric properties of biological tissues:
II. measurements in the frequency range 10 hz to 20 GHz. 41(11):2251.

[34] S. Gabriel, R. W. Lau, and C. Gabriel. The dielectric properties of biological tissues:
III. parametric models for the dielectric spectrum of tissues. 41(11):2271.

[35] S. Gabriel, R. W. Lau, and C. Gabriel. The dielectric properties of biological tissues:
II. Measurements in the frequency range 10 Hz to 20 GHz. Physics in Medicine &
Biology, 41(11):2251, 1996.

[36] S. Godsill and M. Davy. Bayesian harmonic models for musical pitch estimation and
analysis. In 2002 IEEE International Conference on Acoustics, Speech, and Signal
Processing, volume 2, pages II–1769–II–1772.

[37] Andrea Goldsmith. Wireless Communications. Cambridge University Press, Cam-

bridge ; New York, 1 edition edition, August 2005.

[38] S. Guan, J. A. Rice, C. Li, and C. Gu. Automated DC offset calibration strategy for
structural health monitoring based on portable CW radar sensor. 63(12):3111–3118.

88
[39] S. Guan, J. A. Rice, C. Li, and C. Gu. Automated DC offset calibration strategy for
structural health monitoring based on portable CW radar sensor. 63(12):3111–3118.

[40] S. M. A. T. Hosseini and H. Amindavar. UWB radar signal processing in measurement

of heartbeat features. In 2017 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), pages 1004–1007, March 2017.

[41] Carre Technologies inc (Hexoskin). Hexoskin Smart Shirts - Cardiac, Respiratory,
Sleep & Activity Metrics.

[42] Carre Technologies inc (Hexoskin). Hexoskin Smart Shirts - Cardiac, Respiratory,
Sleep & Activity Metrics.

[43] C. Li and J. Lin. Random body movement cancellation in doppler radar vital sign
detection. IEEE Transactions on Microwave Theory and Techniques, 56(12):3143–
3152, Dec 2008.

[44] C. Li, V. M. Lubecke, O. Boric-Lubecke, and J. Lin. A review on recent advances in

doppler radar sensors for noncontact healthcare monitoring. IEEE Transactions on
Microwave Theory and Techniques, 61(5):2046–2060, May 2013.

[45] C. Li, Y. Xiao, and J. Lin. Experiment and spectral analysis of a low-power$ka$-band
heartbeat detector measuring from four sides of a human body. 54(12):4464–4471.

[46] C. Li, Y. Xiao, and J. Lin. Experiment and spectral analysis of a low-powerka-band
heartbeat detector measuring from four sides of a human body. IEEE Transactions
on Microwave Theory and Techniques, 54(12):4464–4471, Dec 2006.

[47] Hongbin Li, Petre Stoica, and Jian Li. Computationally efficient parameter estimation
for harmonic sinusoidal signals. 80(9):1937–1944.

[48] A. Mase, Y. Kogi, and T. Maruyama. Real-time evaluation of heart rate and heart
rate variability using microwave reflectometry. In 2018 IEEE International Microwave
Biomedical Conference (IMBioC), pages 160–162.

[49] S. Nahar, T. Phan, F. Quaiyum, L. Ren, A. E. Fathy, and O. Kilic. An electromagnetic

model of human vital signs detection and its experimental validation. IEEE Journal
on Emerging and Selected Topics in Circuits and Systems, 8(2):338–349, June 2018.

[50] A. Nezirovic, A. G. Yarovoy, and L. P. Ligthart. Signal processing for improved

detection of trapped victims using UWB radar. 48(4):2005–2014.

89
[51] A. Michael Noll. Cepstrum pitch determination. 41(2):293–309.

[52] M. Nosrati, S. Shahsavari, and N. Tavassolian. Multi-target vital-signs monitoring

using a dual-beam hybrid doppler radar. In 2018 IEEE International Microwave
Biomedical Conference (IMBioC), pages 58–60.

[53] M. Nosrati and N. Tavassolian. Effects of antenna characteristics on the performance

of heart rate monitoring radar systems. IEEE Transactions on Antennas and Propa-
gation, 65(6):3296–3301, June 2017.

[54] M. Nosrati and N. Tavassolian. Accuracy enhancement of doppler radar-based heart-

beat rate detection using chest-wall acceleration. In 2018 IEEE International Mi-
crowave Biomedical Conference (IMBioC), pages 139–141, June 2018.

[55] P. Pahl, T. Kayser, M. Pauli, and T. Zwick. Evaluation of a high accuracy range
detection algorithm for FMCW/phase radar systems. In The 7th European Radar
Conference, pages 160–163, September 2010.

[56] B. Park, O. Boric-Lubecke, and V. M. Lubecke. Arctangent demodulation with DC

offset compensation in quadrature doppler radar receiver systems. 55(5):1073–1079.

[57] H. Rohling. Radar CFAR Thresholding in Clutter and Multiple Target Situations.
IEEE Transactions on Aerospace and Electronic Systems, AES-19(4):608–621, July
1983.

[58] F. Schadt, F. Mohr, and M. Holzer. Application of kalman filters as a tool for phase and
frequency demodulation of iq signals. In 2008 IEEE Region 8 International Conference
on Computational Technologies in Electrical and Electronics Engineering, pages 421–
424, July 2008.

[59] K. Thurn, R. Ebelt, and M. Vossiek. Noise in homodyne FMCW radar systems
and its effects on ranging precision. In 2013 IEEE MTT-S International Microwave
Symposium Digest (MTT), pages 1–3.

[60] A. Vallan, M. L. Casalicchio, and G. Perrone. Displacement and acceleration mea-

surements in vibration tests using a fiber optic sensor. 59(5):1389–1396.

[61] ∅. Aardal, Y. Paichard, S. Brovoll, T. Berger, T. S. Lande, and S. E. Hamran. Physical

working principles of medical radar. IEEE Transactions on Biomedical Engineering,
60(4):1142–1149, April 2013.

90
[62] M. Vespe, G. Jones, and C. J. Baker. Lessons for Radar. IEEE Signal Processing
Magazine, 26(1):65–75, January 2009.

[63] Rodrigo Villar, Thomas Beltrame, and Richard L. Hughson. Validation of the Hex-
oskin wearable vest during lying, sitting, standing, and walking activities. Ap-
plied Physiology, Nutrition, and Metabolism = Physiologie Appliquee, Nutrition Et
Metabolisme, 40(10):1019–1024, October 2015.

[64] J. Wang, X. Wang, L. Chen, J. Huangfu, C. Li, and L. Ran. Noncontact distance and
amplitude-independent vibration measurement based on an extended dacm algorithm.
IEEE Transactions on Instrumentation and Measurement, 63(1):145–153, Jan 2014.

[65] J. Wang, X. Wang, Z. Zhu, J. Huangfu, C. Li, and L. Ran. 1-d microwave imaging of
human cardiac motion: An ab-initio investigation. 61(5):2101–2107.

[66] S. Wang, A. Pohl, T. Jaeschke, M. Czaplik, M. Kny, S. Leonhardt, and N. Pohl. A

novel ultra-wideband 80 ghz fmcw radar system for contactless monitoring of vital
signs. In 2015 37th Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC), pages 4978–4981, Aug 2015.

[67] B. Widrow, J. R. Glover, J. M. McCool, J. Kaunitz, C. S. Williams, R. H. Hearn, J. R.

Zeidler, J. Eugene Dong, and R. C. Goodlin. Adaptive noise cancelling: Principles
and applications. 63(12):1692–1716.

[68] P. C. Yuen and G. C. Feng. A novel method for parameter estimation of digital arc.
17(9):929–938.

[69] P. C. Yuen and G. C. Feng. A novel method for parameter estimation of digital arc.
17(9):929–938.

[70] F. Zhu, K. Wang, and K. Wu. A fundamental-and-harmonic dual-frequency doppler

radar system for vital signs detection enabling radar movement self-cancellation. IEEE
Transactions on Microwave Theory and Techniques, pages 1–13, 2018.

91
APPENDICES

92
Appendix A

TI mm-wave FMCW radars

In order to configure the radar, after connecting the radar to a PC, a series of commands
should be sent to the radar for the configuration. The bellow commands are transferred
via UART port to the radar as a command line interface (CLI). Each line is a command.
Some of them mandatory and some of them are optional. But here, we will have a quick
review of the parameters set by these commands and for more details refer to the newest
mm-wave SDK document [4, 2].

sensorStop
flushCfg
dfeDataOutputMode 1
channelCfg 15 3 0
adcCfg 2 1
adcbufCfg -1 0 0 1 0
profileCfg 0 77 7 6 57 0 0 70 1 100 2000 0 0 30
chirpCfg 0 0 0 0 0 0 0 1
frameCfg 0 0 2 0 12.5 1 0
lowPower 0 0
guiMonitor 0 0 0 0 1
vitalSignsCfg 1.5 2.2 256 512 1 0.1 0.05 100000 300000
motionDetection 1 20 2.0 0
sensorStart

The description of the commands are given in the following table:

Table A.1: CLI command description

CLI name Parameters Descriptions

sensorStop Should be the 1st command

93
flushCfg Should be 2nd command
dfeDataOutputMode 1 Frame-based chirp. The advanced
frame configuration is not used here.
channelCfg 15 3 0 4 Rx and 1 Tx are used because this is
a 2D profile.
adcCfg 21 16-bit ADC is used. The ADC output
is complex. It can also be configured
as real. This 1x complex output which
means the image band is filtered out.
Each 16-bit output contains both real
and imaginary parts.
adcbufCfg -1 0 0 1 0 The memory is configured as com-
plex. The position of IQ in the 16-
bit data is determined as I is in LSB
and Q is in MSB. The channel out-
put is not interleaved, which is the only
option supported by AWR16xx and
AWR14xx. The last input is for setting
the threshold for triggering ping/pong
ADC buffer usage. In fact, the ADC
must know when to switch between
ping/pong memories. This is depend-
ing on the number of samples are re-
ceived in a chirp period. The number
of samples depends on the number of
Rx as well. Only option 1 is available
for AWR14xx.

94
profileCfg 0 77 7 6 57 0 0 70 1 100 This configuration can be changed be-
2000 0 0 30 tween stop and start commands. It is
possible to define different chirp pro-
files, though it is possible to define dif-
ferent chirps to put them in a frame. In
AWR14xx only one profile is allowed to
be used in the chirp frame. The start
frequency is set on 77GHz. “idle time
is set 7µs. (float) “ADC valid start
time is set to 6µs. (float) “Ramp end
time = 57µs. This plus the idle time is
equivalent to chirp duration. The “out-
put power back-off is set on 0. (float)
The “tx phase shifter is set to 0. “fre-
quency slope = 70M Hz/µs. This to-
gether with “ramp end time” mean that
the bandwidth is 3.99GHz. (float) “Tx
start time = 1µs. (float) “number of
ADC samples = 100. So, per chirp 100
samples are collected. Probably, this
is managed by the decimating/interpo-
lating filters inside the ADCs. “digi-
tal ADC output sampling rate = 2000
Ksps. The “number of ADC samples =
“ADC sampling rate * “sampling time.
So, “sampling time = 50µs. “High
pass filter1 = has cutoff frequency of
175KHz. “High pass filter2 = has cut-
off frequency of 350KHz. “Rx gain =
30 dB.

95
chirpCfg 00000001 This is a mandatory command. “chirp
start index = 0 “chirp end index = 0
“Profile identifier = 0 “start freq vari-
ation = 0 (Hz) “freq. slope variation
= 0(Hz) “idle time variation = 0 (us)
“ADC time variation = 0 (us) “Tx an-
tenna enabled mask = 1 which means
the first Tx is used. So there is one bit
per Tx antenna showing whether it is
enabled or not. For example, for acti-
vating the 3rd antenna the mask should
be 4.
frameCfg 0 0 2 0 12.5 1 0 “start chirp index = 0 (can be 0-511)
“chirp end index = 0 Because we have
only one chirp profile, the start and the
end indexes are the same. “Number of
loops = 2 (can be 1-255) “Number of
frames = 0 “frame periodicity = 21.5
ms “trigger select = 1 which is the soft-
ware triggered. This is used to deter-
mine how to trigger the FFT units for
starting the process. “frame trigger de-
lay = 0 ms
guiMonitor 0001 detected objects = the export of de-
tected objects is not enabled. “log mag
range is not enabled. “noise profile is no
activated. “range-azimuth heat map is
enabled.
sensorStart Runs the sensor.

96
Appendix B

Matlab codes and the Simulink

model

The phase unwrapping process is written in the listing B.1. The Matlab code for the
optimum filter is written in the listing B.2.

Listing B.1: Phase Unwrapping Matlab code

% The input must contain phases of the ranges where each column is for a
% specific range . In other words , treating each column as an independent
% channel .
%
% UPDATEs :
% September 25 , 2018: after adding or subtracting 2 pi , all the phases to
% the end of the phase vector will be shifted by that amount . Because
each
% shift corresponds to the unit circle shift up or down to a new circle .
% Thus , all later phases should be assumed to be in a new circle .
% Otherwise , the previous procedure did not allow to have a phase beyond
of
% [ -3 pi ,3 pi ].

function UnwrappedPh = unwraphase ( PhaseRange )

% unwrapping the phase

for i = 1: length ( PhaseRange ) -1
cols = PhaseRange ( i +1 ,:) - PhaseRange (i ,:) > pi ;
if ( sum ( cols ) >0)

97
 colsIdx = find ( cols ~=0) ;
for jj = 1: length ( colsIdx )
PhaseRange ( i +1: end , colsIdx ( jj ) ) = PhaseRange ( i +1: end ,
colsIdx ( jj ) ) -2* pi ;
end
end
cols1 = PhaseRange ( i +1 ,:) - PhaseRange (i ,:) < - pi ;
if ( sum ( cols1 ) >0)
colsIdx = find ( cols1 ~=0) ;
for jj = 1: length ( colsIdx )
PhaseRange ( i +1: end , colsIdx ( jj ) ) = PhaseRange ( i +1: end ,
colsIdx ( jj ) ) +2* pi ;
end
end
if cols1 == cols & sum ( cols1 ) > 0
display ( ' At the same time PhaseRange ( i +1) - PhaseRange ( i ) <> \ pi '
)
end

end

UnwrappedPh = PhaseRange ;
end

Listing B.2: Fundamental frequency estimation Matlab code

%% Function of both fundamental frequency estimation and periodic signal
characterization
% [1] M . G . Christensen and A . Jakobsson , O p t i m a l Filter Designs for
% Separating and Enhancing Periodic Signals , IEEE Trans . Signal
% Process . , vol . 58 , no . 12 , pp . 5969 5 9 8 3 , Dec . 2010.
%
% * Inputs * :
% - " y_complex " : the complex input phase
% - " R_hat " : estimated autocorrelation of the input
phase
% - " freqSet " : set of candidate frequencies
% - "M" : averaging length
% - "L" : # of harmonics
% - " pw " : output power vs frequencies in " freqSet
"
% - " p_hat " : the output power of the optimum filter
% - " a_mat " : a matrix with the columns of a
coefficients for each " freqSet " member .

98
% - " h_k_mat " : a matrix with the columns of h
coefficients for each " freqSet " member
%
function [ pw p_hat a_mat h_k_mat ] = f0_a_est ( y_cmplx , R_hat , freqSet ,M , L )

G = length ( y_cmplx ) -M +1;

for ii =1: length ( freqSet )

G_k = zeros (L , M ) ;
W_k = zeros ( L ) ;
for i2 =1: G
w_k_n = exp (1 j *2* pi * freqSet ( ii ) *(1: L ) *( M + i2 -2) ) . ' ;
G_k = G_k + 1/ G * w_k_n *( y_cmplx ( M + i2 -1: -1: i2 ) ) ' ;
W_k = W_k + 1/ G * w_k_n *( w_k_n ) ' ;
end
Q_k = R_hat - G_k ' * W_k ^ -1 * G_k ;
Z_k = MakeA (2* pi * freqSet ( ii ) ,M , L ) ;
h_k_hat = Q_k ^ -1 * Z_k * ( Z_k ' * Q_k ^ -1* Z_k ) ^ -1 * ones (L ,1) ;

pw ( ii ) = ( h_k_hat ) ' * R_hat * h_k_hat ;

p_hat ( ii ) = ones (1 , L ) * ( Z_k ' * Q_k ^ -1* Z_k ) ^ -1 * ones (L ,1) ;
a_mat (: , ii ) = W_k ^ -1* G_k * h_k_hat ;
h_k_mat (: , ii ) = h_k_hat ;
end

end

%% functions added to make matrix A

function A = MakeA ( omega ,M , L )
a_ = @( omega ) exp ( -1 j * omega *(0: M -1) ) . ' ;
A = zeros (M , L ) ;
for jj =1: L
A (: , jj ) = a_ ( omega * jj ) ;
end
end

For the Simulink model, we have shown the main page in Figure 4.17. There are
subsystems, subsubsystems, etc. In Listings B.1 to B.4, the subsystems of the model are
illustrated with their order of placements in the transceiver chain. For instance, Listing
B.1 shows the blocks doing the propagation and collection of the transmitted and received
waves, respectively. In the propagation block, which is the Matlab embedded function, the
transmit array antenna model is formed and it takes the (azimuth,elevation) pair, which is
the angle between the targets and the radar. Similarly, the collector, takes the angle and
computes the received vector by considering the receiver array antenna and its polarization.

99
The polarization is considered, so the output will be two vectors corresponding to the
co/cross-polarizations. We use the co-polarized received signal.
The channel subsystem in Listing B.2, computes the propagation delay and the free-
space pathloss for the travelling wave from the radar platform to the targets location. We
have three targets as explained in Section 4.4, there are three separate channels.
The subsystem of the targets is depicted in Listing B.3 in which a block is associated
to each target. The platform blocks generate vectors of the position and the velocity of
the target at each sample time. The input to each platform block determines the velocity
of the target at a particular time stamp. We assumed that the first two targets have a
velocity proportion to a function of 2πAv fv cos(2πfv t) such that the instantaneous location
of a target is proportion to Av sin(2πfv t). The input velocities realize the vibration of
the targets as intended. Also, in Listing B.3, the MyTarget block reflects the wave at the
targets by the reflection factor related to the RCS of each target.
At the receiver, after downsampling of the beat signal, it goes in the signal processing
subsystem where the frame rate of the received chirps is determined, and then he slow-fast
time matrix is formed. Then, by applying the range FFT, the range-slow time matrix
(map) is obtained. The map goes into the constant false alarm rate (CFAR) block to
detect the range bin containing the reflections of the targets. Then after upsampling, the
phase is calculated and converted to a complex signal in order to find the optimum filters.
We used two independent optimum filter calculations for the two vibrating targets. There
are blocks for demonstrating and saving the necessary signals in the Matlab workspace.

100
Figure B.1: Propagation and collection subsystem

Figure B.2: Channel subsystem

101
 Figure B.3: Targets subsystem

Figure B.4: Signal processing subsystem

102