Application Development on Compact UWB

Radar Systems
Master of Science Thesis [in the Master Degree Programme, Biomedical
Engineering]

QIUCHI JIAN
Department of Signal and Systems
CHALMERS UNIVERSITY OF TECHNOLOGY
Gothenburg, Sweden, 2014
Report No. EX021/2014
Application Development on Compact UWB
Radar Systems

QIUCHI JIAN

Department of Signal and Systems

CHALMERS UNIVERSITY OF TECHNOLOGY

Göteborg, Sweden, 2014

i
Ultra-wide band radar signal processing

© QIUCHI JIAN, 2013

Master’s Thesis EX021/2014

Department of Signal and Systems

Chalmers University of Technology
SE-41296 Göteborg
Sweden

ii
Ultra-wide band radar signal processing
Master’s Thesis in the Master's programme in Biomedical Engineering
QIUCHI JIAN
Department of Signal and Systems
Chalmers University of Technology

Abstract
Ultra-wideband (UWB) radar technique is widely used in various applications from different
domains. The UWB radar system in this work consists of two (Tx and Rx) antennas and one
compact CMOS UWB transceiver. For civilian applications by near field measurements, such
as medical monitoring instruments, parking radars, or ranging and tracking systems, some
desirable properties, including i) compact in size, ii) low cost, iii) fast processing speed, iv)
high spatial resolution and v) high penetration abilities, have been achieved to a large extent
by the system presented in this thesis. This Master project emphasizes two aspects: i) further
investigation on the self-grounded bowtie antenna by using numerical simulations; and ii)
development of signal processing techniques for applications such as breathing and heartbeat
monitoring for medical purposes and sand flow detection for road sanding.

Index Terms: self-grounded bow-tie antennas; UWB radar application; wavelet;

Empirical Mode Decomposition, breathing monitoring; heartbeat detection; sand flow
detection

i
ii
Acknowledgments
I would like to thank everyone involved in this thesis work from both Chalmers University of
Technology and Imego Institute.

To my thesis examiners: Associated Professor Jian Yang, Professor Tomas McKelvey from
Chalmers University of Technology, and Dr. Peter Björkholm from Imego Institute, for their
patience in supervisions and the guidance not only about the thesis work, but also in the
philosophy of scientific research;

To PhD student Yinan Yu from Chalmers University of Technology, for her tremendous help
and continuous encouragement during the work. All this work would not have been possible
without her support;

To Dr. Borys Stoew from Imego Institute, for his kind support. His wide knowledge in many
aspects inspired me deeply. It is a great honor to be instructed by him.

To my colleges in Imego Institute, for the help for daily work. I really appreciate the
memorable time we spent together.

Finally, I would like to thank my parents, without their endless love and selfless support, I
would never be able to come so far in my education. I am forever grateful for what you have
done for me.

iii
List of Abbreviations

API Application Programming Interfaces

CST MWS Computer Simulation Technology Microwave Studio
CR Corner Reflector
EMD Empirical Mode Decomposition
FCC Communications Commission
IMF Instantaneours Mode Functions
MS Magnetic Shielded
Rx Receiving Antenna
SVD Singular Value Decomposition
Tx Transmission Antenna
UWB Ultra-wideband

iv
Table of Contents
1 Introduction .............................................................................................................................. 1
1.1 Ultra-wideband Radar ............................................................................................... 1
1.2 Goal and Tasks .......................................................................................................... 1
2 Self-grounded Bow-tie Antenna Simulation .................................................................. 3
2.1 Simulation Environment and Models ........................................................................ 3
2.2 Results and Analysis.................................................................................................. 4
3 System Calibration .................................................................................................................. 7
3.1 System Structure and Parameters .............................................................................. 7
3.1.1 Antennas ............................................................................................................ 7
3.1.2 Radar Module .................................................................................................... 8
3.1.3 Parameters in the Measurement ......................................................................... 8
3.2 Raw Signal .............................................................................................................. 10
3.3 Temperature Effect .................................................................................................. 11
3.4 Jamming .................................................................................................................. 12
4 Breath Detection .................................................................................................................... 15
4.1 System Setup and Measurement Methodology ....................................................... 15
4.2 Signal Processing..................................................................................................... 16
4.2.1 Signal Pre-processing ...................................................................................... 16
4.2.2 Motion Detection ............................................................................................. 17
4.2.3 Signal Extraction ............................................................................................. 19
4.2.4 Breath Signal Retrieval.................................................................................... 20
4.3 Result and Discussion.............................................................................................. 21
5 Heart Rate Detection ............................................................................................................ 23
5.1 Computer Simulation............................................................................................... 23
5.1.1 System Modeling ............................................................................................. 23
5.1.2 Methodology.................................................................................................... 24
5.2 Experiment with Corner Reflector .......................................................................... 32
5.2.1 Measurement Setup ......................................................................................... 32
5.2.2 Signal Processing............................................................................................. 33
5.2.3 Result and Discussion...................................................................................... 34
5.3 Measurement with Real Human .............................................................................. 36
5.3.1 Motion Detection ............................................................................................. 36
5.3.2 EMD ................................................................................................................ 37
5.3.3 Wavelet Packets............................................................................................... 38

v
 5.3.4 Result and Discussion...................................................................................... 39
6 Sand Flow Detection ............................................................................................................. 42
6.1 Experiment with Funnel .......................................................................................... 42
6.1.1 System Setup and Measurement Methodology ............................................... 42
6.1.2 Results and Discussion .................................................................................... 43
6.2 Measurement with Sand Box................................................................................... 45
6.2.1 System Setup and Measurement Methodology ............................................... 45
6.2.2 Result and Discussion...................................................................................... 45
7 Summary and Future Work................................................................................................ 48
8 References ................................................................................................................................ 50
9 Appendix A: Demo system ..................................................................................................52
9.1 Raw Data ................................................................................................................. 52
9.2 Clutter Removal ...................................................................................................... 53
9.3 One Dimension Tracking ........................................................................................ 54
9.4 Breath Detection ...................................................................................................... 55
9.5 Heart Rate Detection ............................................................................................... 57
10 Appendix B: Paper Published in EuCAP 2014 ............................................................. 59

vi
1 Introduction
Ultra-wideband (UWB) radar technique has been widely used in various applications from different domains. This
thesis work emphasizes the following two aspects:

• Further investigation of the self-grounded bowtie antenna by using a commercial EM (electromagnetic) solver –
CST Microwave Studio;
• Development of Signal processing techniques for the applications such as breathing and heartbeat monitoring for
medical purposes and sand flow detection for road sanding.

1.1 Ultra-wideband Radar

UWB technology has dramatically improved in the past years. In 2002, the Federal Communications Commission
(FCC) formally defined the frequency, bandwidth among other respects with UWB. Besides that, FCC also classified
the possible applications into different categories with constrains, such as data transmission and radar. [1][2]

Several distinct advantages make UWB radar systems attractive for many civilian applications by near field
measurements, such as medical monitoring instruments, parking radars, or ranging and tracking systems [3]. Those
advantages include [2][4]:

• Good time domain resolution that enables the UWB systems to have accurate tracking and positioning;
• Strengthened target recognition;
• Robust to passive jamming;
• Relatively low cost.

1.2 Goal and Tasks

The goal of this project is to develop different prove concept applications with UWB radar systems based on its
advantages. Through this thesis project, the working principle of UWB radar system has been studied and the
possibilities of different UWB radar applications have been developed. Following tasks have been done:

• Computer Simulation Technology Microwave Studio (CST MWS) is used for simulation for self-grounded Bow-
tie antenna with different target surfaces;
• Breath detection;
• Simulations, experiments and real heart rate detection;
• Experiments and real sand flow detection;
• A demo system for possible applications with UWB radar system.

1
2
2 Self-grounded Bow-tie Antenna Simulation
In order to understand the behaviour of the self-grounded bow-tie antenna which is a compact UWB antenna with
many applications [5]-[10], a set of models for different scenarios of a self-grounded bow-tie antenna are simulated
and analysed using CST MWS.

2.1 Simulation Environment and Models

Six models, referred to as models 1 to 6, with different setups have been run successfully with 10ns steady time. A
self-grounded bow-tie antenna is used in all models, and, as shown in Table 1, each model has its own environment
setup in order to simulate different measurements: model 1 has only the antenna; models 2 and 3 have an iron surface
placed toward the antenna with different distances; models 4 to 6 have different surface conditions: surface covered
by water layer, water drops and ice layer. Figure 2-1 is the screen shot of one of those simulation models.

Table 1 Setup of simulation models

Model # 1 2 3 4 5 6
Antenna Size
Face to an Iron Surface No Yes
Surface Size N/A x:-500~500; y: -500~500; z: -10~0.
Surface Property N/A Electric Conductivity: 1.04e+007; Mue: 1.
Dis. from Antenna(mm) N/A 300 600 300
Covered by N/A Water Water Drops Ice

Figure 2-1 Screen shot of a simulation model

3
2.2 Results and Analysis

Figure 2-2 Results from model 1, 2 and 3 in time domain

Figure 2-2 contains the results from the first 3 models. The curve in black is from the model with a self-grounded
bow-tie antenna only. The dotted lines in red and green are from models with the distance 300mm and 600mm
between the antenna and the iron surface, respectively. It can be seen from the graph that all 3 curves coincide from
the start to about 3.5 ns. After that, the red and the green dotted lines deviate from the black curve at around 3.8 and
5.8 ns, respectively.

In order to remove the interference from the antenna itself, the signal from model 1 is used as a clutter map. In Figure
2-3, the red curve and the blue dotted curve are made by subtracting the clutter map from the signal from model 2 and
model 3, respectively. By the setup of these simulation models, it is clear that the reflection should be received after
time tr, which could be calculated by

𝑡𝑟 = 𝑡𝑎 + 𝑡𝑖𝑛𝑐𝑖𝑑𝑒𝑛𝑡 + 𝑡𝑟𝑒𝑓𝑙𝑒𝑐𝑡𝑖𝑜𝑛 (2.1)

where ta is the time the signal traveling inside the antenna (cable is included), tincident and treflection are the time that the
signal travels to and from the iron surface. Considering there is only one antenna in our setup and all objects remain
still, the tincident and treflection should always be the same, and the traveling time in the air could be calculated by traveling
distance over 3 * 1011 mm/s, the equation could reduce to
𝑑𝑡𝑎𝑟𝑔𝑒𝑡
𝑡𝑟 = 𝑡𝑎 + 2 ∗ 3∗1011 𝑚𝑚/𝑠
(2.2)

4
Therefore, the tr of the red curve is
300 𝑚𝑚
𝑡𝑟𝑟𝑒𝑑 = 𝑡𝑎 + 2 ∗ 3∗1011 𝑚𝑚/𝑠 = 𝑡𝑎 + 2 𝑛𝑠 (2.3)

And the tr of the green curve is

600 𝑚𝑚
𝑡𝑟𝑏𝑙𝑢𝑒 = 𝑡𝑎 + 2 ∗ 3∗1011 𝑚𝑚/𝑠 = 𝑡𝑎 + 4 𝑛𝑠 (2.4)

Thus, the difference of the reflection time between the blue and the red curve should be 2ns. As we can see from
Figure 2-3, the distance between the main peak of the red curve and the blue one is around 2 ns, which concurs our
theoretical calculation. Besides that, a significant attenuation of signal strength with the increasing distance between
the antenna and the target can also be observed from the figure.

Figure 2-3 Clutter removed results from model 2 and 3 in time domain

Figure 2-4 contains the results in frequency domain from all models but model 2. The reference curve here is the
black curve from model 1, which contains only the antenna. The rest models have both the antenna and a surface
placed 300mm from the antenna with different conditions, such as iron surface, water layer covered iron surface and
so on. Based on the shape of these curves, we could divide them into two groups, where group A contains model 3
and 4, and group B contains model 5 and 6.

The two curves in group A share some features with the black curve on the shape, but with significant magnitude
differences especially in the range of 0.5 to 1.8 GHz. These differences could be useful for distinguishing surface
conditions. On the other hand, the curves from group B are quite different from the reference curve. Therefore, it
should be possible to make a distinction between these two conditions and the model 1. Considering the temperature
could be good reference information, although these two curves share high similarity with each other, it is possible to
determine the surface conditions.

5
Figure 2-4 Results of all models except model 2 in frequency domain

6
3 System Calibration
System calibration is applied to clarify the system property and identify potential interferences that could compromise
the measurements, such as temperature effect and jamming. Adjustments corresponding to what have been done on
the further experiment environment setups.

3.1 System Structure and Parameters

Figure 3-1 Structure diagram of the radar system

Figure 3-1 shows the structure of the radar system that is used in this work. The system is mainly composed of a
computer, a Novelda transceiver, a transmission antenna (Tx) and a receiving antenna (Rx). The antennas are
connected to the transceiver through SMA (subMiniature version A) connectors and the transceiver is connected to
the computer via a USB cable. Signal and transceiver parameter acquisition and update is done through Application
Programming Interfaces (APIs) that are provided by Novelda’s library Radarlib3. [RadarUserGide]

3.1.1 Antennas

Antennas used in this project are mainly Vivaldi 004 and Vivaldi 006, as shown in Figure 3-2. Table 2 contains some
properties of these antennas:

Table 2 Main parameters of antennas [11][12]

Dimensions Frequency band

Vivaldi 004 150mmx133mmx1.6 mm 900 MHz – 5 GHz
Vivaldi 006 52mmx52mmx1.52 mm 3 – 20 GHz

Figure 3-2 Photos of antennas. From left to right: the front side of the Vivaldi 004 antenna; the back side of the
Vivaldi 004 antenna; the front side of the Vivaldi 006 antenna; and the back side of the Vivaldi 006 antenna.[11][12]

7
3.1.2 Radar Module

Figure 3-3 Photo of the Novelda radar module

Novelda radar module contains a NBA6100 chip with transmit bandwidth from 450MHz to 3.55GHz, a flash for
storing data, a USB2.0 IO module as power supplier as well as data communicator to computer with a data rate of
480Mb/s speed [15].

3.1.3 Parameters in the Measurement

Table 3 contains some parameters that are used in the measurements in this work.

Table 3 Novelda radar system parameters [15]

Parameter Name Description Value

Averaging Factor The number of internal samples 100
(AF) is taken before output, which
affects the processing speed and
the noise level.
Gain The gain of the transmitted 5
signal.
Frame Offset (FO) Internal time delay (including 290(Vivaldi004)
both antennas and cables): 278(Vivaldi006)
decides the range of the view for
the radar.

8
Pulse Repetition The frequency the pulses are 48 MHz
Frequency (PRF) sent.
DAC Digital to analog Converter. 8192
Convert the signal to 13 bits.
Distance resolution Distance between each sampler. 3.7137 mm (in air)
(DS)
Samplers Per Frame The number of samplers in each 512
(SPF) frame.
Range The maximum distance the radar From the outer edge of the
could detect, with the FO setting, antenna extends to about 1.9
the range could be calculated m (in air)
by Range = DS ∗ SPF.
Frame Rate (FR) The number of frames that could ≈ 59 frames/s theoretically,
obtained per second, could be the real frame rate depends
calculated by FR = on the condition of the
𝑃𝑅𝐹 ∗ DAC1∗ AF . computer.

9
3.2 Raw Signal

The NVA6100 radar module supports sampling window with 512 parallel samplers [15]. Therefore, data from each
sampling window is a 512 elements vector, which is called as a frame of signal. As shown in Figure 3-4, a frame of
signal collected at time t is represented by a vector r, and each element in the vector, which is represented by r(i), is
the data from sampler i at time t. The frame’s view can descript the environment regarding the distance from the
antenna at certain time t.

Figure 3-4 An example of a frame of data that is collected at time t

By collecting signals over time, a signal matrix X, as shown in Figure 3-5, can be constructed. In the figure, the
sampling time is represented by from t1 to tn. The data X (t, s) represents the data collected by sampler s at time t.
Besides the frame’s view, with a data matrix, it is also possible to observe the changes on a specific sampling point
among time, which is called sampler’s view. Different from the frame’s view, which can reflect the changes in certain
range, the sampler’s view focus more on the changes at a fixed position. Based on the different properties of these two
views, a more comprehensive observation can be obtained.

Figure 3-5 An example of a data matrix that is collected from time t1 to tn

10
3.3 Temperature Effect

During the system calibration, by looking into the sampler’s view of raw signal, it is observed that there is a heating
effect in the system. This could due to the temperature of the chip raises rapidly in the first period of the measurement.
In order to determine the time it takes to reach a stable level, several measurements are done.

All measurements are done in a still environment with heating time from 0s to 900s. The results from three
measurements with start-up time 0, 90 and 180 seconds are shown in Figure 3-6. It can be seen from the figure that in
the first 60 seconds, the signals from each sampler affect by the temperature effect significantly. However, after
around 90 seconds, the signal could get over the temperature effect.

Figure 3-6 The raw signal from several samplers with different start-up times in the sampler’s view.

11
3.4 Jamming

Figure 3-7 The raw signal from the 20th sampler during 5 seconds period.

From the analysis of measurements, it is clear that there is some unknown interference with certain pattern in the
signal. Some more data were collected in the same place with no moving object.

The Figure 3-7 above shows the signal from the twentieth sampler in the frame for five seconds. As we can see, there
are some pluses with frequency around 2 Hz. Considering in this study, the signals with lower frequencies are the
main interest, such interference might significantly affect our observation. Therefore, it is important to find out the
possible interference source.

Above all, the power frequency could be ruled out. This is because not only that the plot shows much lower frequency,
but also because the sampling rate is around 60 Hz, which is far below the minimum sampling frequency that defined
by Nyquist–Shannon sampling theorem.

The noise is highly possible emitted from some device around, considering it has a very strong pattern. Thus, several
measurements were done for verification with the setup in Table 4.

First of all, in order to exclude electromagnetic interference from the cables connecting the antennas, transceiver and
computer, a shielding ring was used in different places for several measurements. The signal is affected very little by
using a shielding ring on single wire between the transceiver and the antenna, or using it on both wires at the same
time. Thus, the cables are not the major source.

Beside cables, the transceiver itself could be a source. By changing the sampling rate and PRF, the transceiver did
sampling and generated pulses in different frequencies. However, the interference remained the same. Therefore, the
transceiver should not be the source.

12
 Table 4 Experiment setup for determining jamming source

Possible Control Group Setup Experiment Setup Comparison result

Source
SMA cables No shielding ring at all Shielding ring on Tx Have slightly
connecting to cable difference
antennas No shielding ring at all Shielding ring on Rx Have slightly
cable difference
No shielding ring at all Shielding ring on both Hardly see any
cables at the same time differences.
USB cables No shielding ring at all Shielding ring on the Hardly see any
connecting to cable differences.
computer
Radar Module Sampling rate set to 0 Sampling rate set to 1 Jamming is still there
Sampling rate set to 0 Sampling rate set to 2 Jamming is still there
PRF set to 48 MHz PRF set to 6 MHz Jamming is still there
PRF set to 48 MHz PRF set to 12 MHz Jamming is still there
PRF set to 48 MHz PRF set to 24 MHz Jamming is still there
Desktop System connected to a System connect to a Jamming is still there
Computer desktop computer laptop with plugin
power source
System connect to a System connect to a Jamming is still there
laptop laptop with battery as
power source

The next possible interference source device was the desktop computer. In order to eliminate the effect from it, a
laptop computer with battery as power source was used for replacement. However, the interference still remained.

Considering all electronic devices in the room are excluded, the only possibility left is that the interference is from
electromagnetic field generated somewhere else. Therefore, the radar system was moved to a magnetic shielded room
and connected to a computer outside the room with the cable through a small hole on the wall. Following is the
comparison of the result and it confirmed that the interference source is not from the system itself.

13
Figure 3-8 Raw signals from 3 samplers (20th, 24th, 28th) in a still environment in a magnetic shielded room and a
normal office.

14
4 Breath Detection
4.1 System Setup and Measurement Methodology

Figure 4-1 Diagram of the experimental setup of the breath detection using the present UWB radar system. The graph
on the left hand side shows the setup from the side view and the right hand side one shows from the top view.

Figure 4-1is the setup diagram for breathing detection: two Vivaldi 004 antennas Tx and Rx are placed in parallel with
12cm distance between each other, which is represented by da in the figure. One person is sitting still in a chair with a
distance 24 cm, which is represented by dobj in the figure, between the chest and the outer edge of the antennas. In the
experiment phase, measurements have been done for 60 seconds continuously. From previous test, the frame offset
should be 290 in order to set the detection range from the outer edge of the antennas.

15
4.2 Signal Processing

Figure 4-2 Signal processing procedure diagram for breath detection.

The Figure 4-2 is the flow chart of the signal processing procedures. The collected raw signal first passes the pre-
processing procedure to extract signal from the target area, where contains the most interesting signal, and remove the
clutter. After that, the signal is used for two purposes, one is to verify the target’s presence, and the other is for
retrieving the breath signal.

4.2.1 Signal Pre-processing

Considering the radar has a relatively wide range compared to the size of human’s chest, only partial signal from the
target area is taken for further process, which based on the distance between human and the antennas, for time gating
the signal.

For example, the Figure 4-3 shows a frame of signal. During the measurement, the person was sitting about 240mm
from the outer edge of the antennas. By the radar’s distance resolution, which is 3.7137 mm, we deduce that the
person should be in the area after sampler 65. Considering the human in the measurements has about 180 mm thick
chest and taking possible measurement errors into account, target area was set to between sampler 55 and sampler 104
as showed in the figure in red.

After obtaining the data of interest, the clutter removal is followed. The previous developed singular value
decomposition (SVD) method for clutter mapping [4] was used here. The outcome data was used for further
processing.

16
 Figure 4-3 A frame of signal collected when there is a person in its detection range.

4.2.2 Motion Detection

In order to confirm the received signal contains information about any moving object, a differential phase method is
applied here for motion detection. For two identical signals with a time delay, the time delay could be obtained by
Fourier Transform. [4]

Suppose s1 and s2 are identical signals with a phase shift:

𝑠2 (𝑡 ) = 𝑠1 (𝑡 − 𝜏) (4.1)

By applying Fourier Transform we have:

𝑆1(𝜔) = 𝑆2(𝜔) 𝑒 −𝑗𝜔𝜏 (4.2)

Therefore:

𝑆1(𝜔)
= 𝑒 −𝑗𝜔𝜏 (4.3)
𝑆2(𝜔)

Thus, the angle difference could be obtained by:

𝑆1(𝜔)
∠ = −𝜔𝜏 (4.4)
𝑆2(𝜔)

Considering human breath is a slow periodic movement and the radar has a high sampling rate, in most of the samples,
the target has either small movement or remain still comparing with the previous sample. Therefore, the distribution
of the phase shift should have a very high peak around 0 and then drop dramatically on both sides. On the other hand,

17
in an environment with random noise, if there is nothing moving periodically in the detection range, the curve
calculated with the measurement data would show a Gaussian distribution. Therefore, this could help us to detect the
presence of target.

Figure 4-4 shows an example of motion detection. Two measurements were done with almost the same environment
setup. The only difference is in the second measurement, there was a person sitting in the target area breath normally.
The blue curve is calculated from the measurement data with no person presence and the red on is from the
measurement data that one person was sitting in the range. As we can see from the figure, these two curves can be
easily distinguished by both the peak value and the shape.

Figure 4-4 Calculation results based on the data collected with and without a person present in the radar’s detection
range.

18
4.2.3 Signal Extraction

Figure 4-5 Result after applied Fourier Transform on pre-processed signal.

The continuous phase could be extracted by applying Fourier Transform on the pre-processed signal. By finding the
coefficients from all samplers in each frame, the curve in Figure 4-5 could be obtained.

Assume the normal breath rate is in the range of 6 to 40 times per minute, a band-pass filter with bandwidth from 0.1
to 0.7 Hz is applied. Figure 4-6 shows the design view of the filter.

Figure 4-6 Design view of the band-pass filter.

19
4.2.4 Breath Signal Retrieval

In order to avoid counting small false peaks into the result, the peak counting procedure follows the algorithm below:

Calculate the mean value mean and the standard deviation std of the processed signal.

Set the possible range of the distance between two true peaks min and max based on the fact that breath frequency that
we are interested in, which is from 0.1 to 0.7 Hz with the formulas below:
frameRate
𝑚𝑖𝑛 = 0.7
(4.5)

frameRate
𝑚𝑎𝑥 = 0.1
(4.6)

Find all peaks on the breath curve and save in a list p. Go through elements in p in order with following procedure:

• Create an empty vector peaks for breath peak storage;

• For each element pi:
if 𝑝𝑖 > 𝑚𝑒𝑎𝑛 + 𝑠𝑡𝑑 𝐴𝑁𝐷 𝑓𝑟𝑜𝑚 𝑙𝑎𝑠𝑡 𝑏𝑟𝑒𝑎𝑡ℎ 𝑝𝑒𝑎𝑘 < 𝑚𝑖𝑛:
if 𝑝𝑖 < 𝑝𝑖−1, pi is NOT a breath peak;
else pi-1 is NOT a breath peak, pi is a breath peak;
else if 𝑝𝑖 > 𝑚𝑒𝑎𝑛 + 𝑠𝑡𝑑, pi is a breath peak;
else if mean + std > 𝑝𝑖 > 𝑚𝑒𝑎𝑛 𝐴𝑁𝐷 𝑓𝑟𝑜𝑚 𝑙𝑎𝑠𝑡 𝑏𝑟𝑒𝑎𝑡ℎ 𝑝𝑒𝑎𝑘 ≥ 𝑚𝑎𝑥, pi is a breath peak;
else pi is not a breath peak.

At the end, count the number of breath peaks.

20
4.3 Result and Discussion

Figure 4-7 shows the frequency domain of one sample among all measurements before and after the band pass filter,
which are in black and blue, respectively. As shown in the figure, the signal with higher frequency has been
sufficiently suppressed.

21
 Figure 4-7 Signal before and after the band-pass filter in frequency domain.

Figure 4-8 Result of a one minute sample after peak counting.

Figure 4-8 shows the final result from a one minute sample after peak counting. The blue curve indicated the signal,
while the red and green lines are the mean value and the standard deviation value. The peaks with pink point are those
we counted as a breath peak, which are 15 in total. The actual breath was 14 times during the measurement. The result
is relatively good.

22
5 Heart Rate Detection
5.1 Computer Simulation
Before real experiment, a computer simulation has been done. Three methods were used to process the simulation
signal, which were band-pass FIR1 filter, Empirical Mode Decomposition (EMD) and Wavelet packets.

5.1.1 System Modeling

We assume that despite, the signal from a still human contains information mainly about the breath and the heartbeat
even though there are random noises and movements from other inner parts of human body, such as blood flow.
Moreover, it is assume that the breath and the heartbeat of a human in a short period of time, such as one minute,
remain stable with regular frequencies. Based on these assumptions, two sine waves with different amplitude and
frequency were used to simulate human heartbeat (the black curve in the figure below) and breathe (the blue curve),
respectively. In this model, the simulated heart rate is 90 times per minute (1.5 Hz) and the breath rate is 15 times per
minute (0.25 Hz). A random number 0.956 is used as the initial phase difference. The composed signal (the red curve)
from these two sine waves was used to simulate the measured signal. All three signals were generated by:

heartBeat(t) = 0.5 ∗ sin(2 ∗ pi ∗ 1.5 ∗ t) (5.1)

breath(t) = 4 ∗ sin(2 ∗ pi ∗ 0.25 ∗ t + 0.956) (5.2)
composedSignal=heartBeat(t) + breath(t) (5.3)

Figure 5-1 Heart rate simulation signal and its components.

Figure 5-2 shows the composed signal in the frequency domain. There are two large peaks that indicate the
frequencies of breath and heartbeat, which are around 0.25 and 1.5 Hz respectively.

23
 Figure 5-2 Simulation signal in frequency domain.

5.1.2 Methodology
5.1.2.1 Band-pass Filter

For this model, a FIR1 band-pass filter is naturally used for the first attempt to retrieve the heartbeat signal. Figure 5-3
shows the filtered signal in frequency domain. It can be seen that although the filter is able to depress the breath signal
significantly, there is still a small peak at around 0.25 Hz, which indicate that might not be sufficient for this problem
by using a single filter.

Figure 5-3 Simulated signals and reconstructed heart beat signal in frequency domain.

24
The figure below compares the processed signal and the original simulation signal. As we can see, the processed
signal basically keeps the same shape with the original one, but with a little shift on the time and some amplitude
differences on peaks.

Figure 5-4 Simulated heart beat signal and the filtered reconstructed heart beat signal.

5.1.2.2 Empirical Mode Decomposition (EMD)

Empirical Mode Decomposition (EMD) is a nonlinear technique that can decompose a signal into finite compnonets
known as Instantaneours Mode Functions (IMF), which was developed by N.E. Huang in 1998. [13]

The algorithm works as below: for a signal s(t), which is the black curve in Figure 5-5: [13]

• Find all minima and maxima extremes and plot a curve for each set by using cubic spline interpolation, which are
the blue and red curves respectively.
• Compute the mean values between the maxima and minima curves, and make a residual curve r(t) which is the
pink one in the figure.
• Get the detail known as IMF by d(t) = s(t) − m(t).
• Repeat the process above on the residual r(t) until it fits the stopping criteria.

The stopping criteria can be defined based on the requirement of the analysis. In this thesis project, a existing library
is used for calculation which defines the calculation stops once all condition below are satisfied:

• For each point, meanAmplitude(t) < 0.5* amplitude(t);

• mean of boolcriteria(t) less than 0.05;
• There are less than two extremes in the curve.

where
max (t) + min (𝑡)
meanAmplitude (t) = abs (Amptude Amplitude
2
) (5.4)

25
 max (t) − min (𝑡)
Amplitude Amplitude
amplitude(t) = abs ( 2
) (5.5)
mean (t)
Amptude
boolcriteria (t) = (amplitude(𝑡) ) > 0.05 (5.6)

By this algorithm, the given signal is expected to be decomposed into several sub signals with different frequencies.

Figure 5-6 shows the outcome of the simulated signal by applying EMD. As showed in the figure below, the
decompose procedure stops after the second iteration. The graph on the top is the composed simulation signal over 50
seconds time period. The two graphs below are the IMFs from the first and second iteration, which are referred as
IMF1 and IMF2 respectively.

IMF1 is a sine-like wave with 75 peaks during 50 seconds (1.5 Hz) and average amplitude of about 0.5, which fit the
original heartbeat simulation signal. IMF2 is also a sine-like wave, with 13 peaks during 50 seconds (0.26 Hz) and
average amplitude of about 4. Considering the last peak is from an uncompleted cycle, it is reasonable to ignore the
slightly difference between IMF2 and the original breath simulation signal. Therefore, both the heartbeat and breath
signals are nicely retrieved by EMD.

Figure 5-5 Signal and residue after the first iteration with EMD method.

26
 Figure 5-6 IMF1 and IMF2 of the composed simulation signal.

Figure 5-7 shows the composed signal, the heartbeat simulation and the IMF1 in frequency domain. It is clear that by
using EMD, the signal in lower frequency range was suppressed sufficiently while the heartbeat signal was remain
almost untouched.

Figure 5-7 Simulated signals and reconstructed heart beat signal in frequency domain.

27
From comparing the simulation signal and IMF1 in time domain, it can be seen in Figure 5-8 that after applying the
EMD method, the heartbeat signal could be nicely retrieved.

Figure 5-8 Simulated heart beat signal and the filtered reconstructed heart beat signal.

5.1.2.3 Wavelet Packets

Wavelet is a technique for decompose time series with different scales and times. [8] By using Wavelet transform, a
signal could be divided into two parts: Signals in lower frequency (A, approximation coefficient) and higher
frequency (D, details). By using wavelet packets, the Wavelet transform is first applied to the original signal and get
A and D. Then the transform is applied again on both A and D and their outcome, respectively.

In other words, this analysis will produce a binary tree like in Figure 5-9. S is the original signal, which is the root
node (0, 0) on the top of the tree. By applying wavelet transform on node (0, 0), signal could be decomposed into two
parts, node (1, 0) and node (1, 1), which are the sub signal in lower frequency and the higher frequency respectively in
level 1. The decompose procedure could continue with both nodes to get to the 2nd level of the tree with 4 nodes.
Therefore, on the nth level, the tree should have 2n nodes. Considering the signal could be reconstruct from any node,
this method could enable us to select the signal in the frequency that is interested in.

28
 Figure 5-9 Binary tree produced by wavelet packets methods.

Figure 5-10 the composed simulation signal and two signals with target frequencies.

The red signal in Figure 5-10 is the composed simulation signal. The rest two blue curves are from the node (5, 1) and
node (6, 1) with the target frequencies. The reconstruction signal was composed by these two signals.

From Figure 5-11 it can be seen clearly that the reconstructed signal perfectly fits the heartbeat simulation signal in
the frequency domain and the breath signal is nicely removed from it.

29
 Figure 5-11 signals and reconstructed heart beat signal in frequency domain.

Figure 5-12 Simulated heart beat signal and the filtered reconstructed heart beat signal.

The red curve in Figure 5-12 is the reconstructed signal in time domain, composed by signals from the two nodes and
the black curve is the simulation signal. It is obvious that the reconstruction signal basically fits the simulation signal
with rather small errors.

30
31
5.2 Experiment with Corner Reflector
5.2.1 Measurement Setup

In order to quantify the detection capability of the radar system for small movements (±0.5mm), such as human
heartbeat, an experiment has been performed.

The experiment system contains following components:

 An Impo 1417 type function generator;

 An Impo 1419 type vibrator;
 A corner reflector(CR) which was made of three 6cm*6cm*6cm size boards with copper surface;
 The Novelda transceiver with two Vivaldi 004 antennas.

To rule out jamming from other sources, the measurement was performed in a magnetic shielded (MS) room. The
whole test system except the desktop computer was placed inside the MS room. In Figure 5-13, the transceiver was
hanged behind the wooden bracket, which could not be seen from this angle.

Figure 5-13 Photo of the experimental setup of corner reflector movement detection using the presented UWB radar
system.

As shown in Figure 5-14, two Vivaldi 004 antennas were placed in parallel with 17cm distance. The vibrator and
corner reflector were placed about 27cm from the outer edge of the antenna in vertical and right in the middle of two
antennas. Therefore, the actual distance from the object to antennas is around 28cm.

32
 Figure 5-14 Diagram of the experimental setup of corner reflector movement detection using the presented UWB
radar system.

Samples were taken with parameters in Table 4. Both the signal and the detection time were saved as matrix and
vectors respectively.

Table 5 Experiment parameters.

Measuring dobj da CR Excitation Excitation

Time movement Frequency Waveform
≈60s ≈27cm ≈17cm ±0.5mm 1,1.5,2,2.5,3,3.5 Hz Sine, Triangular

5.2.2 Signal Processing

The signal processing procedure followed the diagram in Figure 5-15. All three methods, SVD, EMD and Wavelet
Packets, are applied to the experimental signals separately. After signal has been processed, a peak counting method is
applied to get the final result. In order to avoid counting small false peaks into the result, the peak counting method is
designed like below:

• Calculate the mean value (mean) and the standard deviation (std) of the processed signal.
• Apply the threshold below to the signal:

mean − std, 𝑠𝑖𝑔 < 𝑚𝑒𝑎𝑛

f(sig) = {
mean + std, 𝑠𝑖𝑔 ≥ 𝑚𝑒𝑎𝑛

• Count the number of points in the signal after threshold that fit the condition: f(s) = mean − std AND f(s +
1) = mean + std.

33
 Figure 5-15 Signal processing procedure diagram for corner reflector movement detection.

5.2.3 Result and Discussion

Figure 5-16 the results of signal of one sample with 1.5 Hz movement processed in the frequency domain.

Figure 5-16 shows the result in the frequency domain. The black curve is the preprocessed signal. The signal after the
filter, EMD and Wavelet packets are in blue, red and green. It is obvious that the signal after filter contains relatively
more signal around the target frequency. Considering in reality, the breath signal is much stronger than heartbeat
signal and their main frequecie are very close, the FIR1 band-pass filter method can hardly work.

34
All three methods have the same result of 91 beats for one minute (1.52Hz). This is quite close to the simulation
frequency 1.5 Hz, which should be 90 beats ideally. However, the EMD has a relatively large error on signal
amplitude as shown in Figure 5-17, which might be caused by frequency leakage. Frequency leakage here refers the
phenomenon that IMF contains the signal which does not belong to its fundamental frequency. This could be caused
by several reasons. One of them is that the mean values are calculated by interpolation values based on the minima
and maxima extremes sets, which means they are not the true mean values. Moreover, in each iteration, the IMF is
only a partial signal from the original one that is highly effected by the calculated mean values. Therefore, all
approximations in the calculation make frequency leakage unavoidable in EMD method. [17] Besides that, due to the
difficulty to pre-determine the fundamental frequency of each IMF, the reconstruction work could be tricky because it
is hard to predict which IMF should be included into the result.

Figure 5-17 Results of different processing method applied on the experiment signal.

35
5.3 Measurement with Real Human

The measurement setup remained the same with the breath detection. The raw signal was processed with following
procedure shown in Figure 5-18.

Figure 5-18 Signal processing procedure diagram for heartbeat detection

5.3.1 Motion Detection

Figure 5-19 shows the motion detection result from a one-minute sample with 78 times heartbeat. With the presence
of a human, the peak raised from around 700 to 2300. This indicated that a strong periodic signal is obtained.

36
 Figure 5-19 Motion detection result of one sample with one human sitting in the radar range.

5.3.2 EMD

In the Figure 5-20, the red curve is the preprocessed signal. Following four curves are the signals from IMF 3, 4, 5
and 6 that are used to reconstruct the signal, which is the curve at the bottom.

Figure 5-20 IMFs produced by applying EMD on pre-processed signal and the signal composed by them.

37
The outcome after filtering is shown in Figure 5-21. There are 75 beats counted from the processed signal, which has
3 beats missing. By using a simulation sine wave with frequency 1.2 Hz as a reference, it shows that some very small
peaks were seen as one peak with other peaks.

Figure 5-21 the reconstructed signal after filtering and peak counting.

5.3.3 Wavelet Packets

Figure 5-22 Signals from different nodes that are produced by applying wavelet packet on pre-processed signal and
the signal composed by them.

38
In the Figure 5-22, the red curve is the preprocessed signal. Following three curves are the signals from node (4, 1), (5,
1), and (7, 1) that are in the human heart rate frequency range, which is from around 0.7 to 3.75 Hz. The curve at
bottom is the composed signal from the three nodes, which is sent to the filter before peak counting.

The outcome after filtering is shown in Figure 5-23. There are 77 beats counted from the processed signal, which is
very close to the real heartbeat 78 times. By using a simulation sine wave with frequency 1.2 Hz as a reference, it
shows that the reconstructed signal basically match the shape of the simulation.

Figure 5-23 the reconstructed signal after filtering and peak counting.

5.3.4 Result and Discussion

Figure 5-24 shows the results in frequency domain. The black curve is from the preprocessed signal and the blue and
red are from the results after EMD and Wavelet Packets. It is clear that the preprocessed signal contains many large
peaks in the frequency range from 0 to 1 Hz, which mainly came from the breath. By applying EMD and Wavelet
packets methods, not only signal in this range has been sufficiently depressed, but also the signal in higher frequency.
As we can see from the figure, main peaks after process concentrate in the range 1.1 to 1.4 Hz. Considering the
frequency spectrum has overall information of the sample for one minute and the real heartbeat is with around 1.2 Hz
frequency, the results are relatively good.

39
 Figure 5-24 Pre-processed signal and the signal extracted by wavelet after filter in frequency domain.

Figure 5-25 shows the reconstruction result in time domain. Considering two curves share high similarity on their
shapes and the number of peaks, it should be able to say that the heartbeat signal from a real human could be obtained
by the radar system and be extracted with different signal processing methods.

Figure 5-25 the heartbeats are reconstructed by sinusoid signals. This figure shows the normalized reconstructions
from continuous measurements. The accuracy is evaluated by comparing the number of real heartbeats and
reconstruction peaks.

40
41
6 Sand Flow Detection
Another interesting attempt of UWB radar application is sand flow detection. In icy winter days, sand is widely used
for increasing the friction between the road surface and wheels. Trams normally carry a sandbox that can provide the
sand in need. However, because the sand is pumped under the tram, the driver cannot really know if the sand has
actually been pumping out. Therefore, it would be very interesting if a UWB radar application could do the
monitoring work.

6.1 Experiment with Funnel

The main idea of the application is to see if the UWB radar can detect the sand flow in the range. In order to get rid of
the noise of complex environment, which is the tram in this case, an experiment is done.

6.1.1 System Setup and Measurement Methodology

Two kinds of measurements were performed for tring detect the reflection and transmission of the sand flow.

Figure 6-1 shows the digram of the transmission detection expirment. Two Vivaldi 004 antennas are placed at both
side of the sand flow with the distance dobj = 54 cm. The distance between antennas and the bottom of the funnel h = 7
cm.

Figure 6-1 the diagram of the experimental setup for detecting the transmission of the sand flow. Graph on the left
hand side shows the setup from the front view and the one on the right hand side shows from the top view.

Figure 6-2 shows the digram of the reflection detection expirment. Two Vivaldi 004 antennas are placed at the same
side with the distance da = 10 cm in between. Both of them face to the sand flow with the distance from the outer edge
dobj = 25 cm. The distance between antennas and the bottom of the funnel h = 8 cm.

42
Figure 6-2 the diagram of the experimental setup for detecting the reflection of the sand flow. Graph on the left hand
side shows the setup from the front view and the one on the right hand side shows from the top view.

6.1.2 Results and Discussion

The result in transmission detection is shown in Figure 6-3. The blue line is the mean value of all frames that collected
with no sand flow. The red line is the mean value of all frames that mesured with sand falls out of the sand by gravity.
It can be seen clearly that the signal strength is reduced when the sand flow is passing the radar’s range.

Figure 6-3 Measurement results for transmission detection of the sand flow.

The result in reflection detection is shown in Figure 6-4. The blue line is the mean value of all frames that collected
with no sand flow. The red line is the mean value of all frames that mesured with sand falling. As we can see, there is
hardly any difference between these two signals, which indicate the sand flow has almost no reflection at all.

43
 Figure 6-4 Measurement results for reflection detection of the sand flow.

This phenomenon could be explained by the form of the sand flow. Unlike metal plate, which has a flat surface that
could reflect the incident waves directly back to the source, the sand flow scatters the incidences to arbitrary
directions, and therefore the detectable reflections are not visible to the radar system.

44
6.2 Measurement with Sand Box
6.2.1 System Setup and Measurement Methodology

Figure 6-5 the diagram of the experimental setup for detecting the transmission of the sand flow from a real sand box.
Graph on the top shows the setup from the front view and the one on the bottom shows from the top view.

Considering the transmission measurement has a better result in the previous experiment, it is used to measure the real
sandbox that is used on trams. Figure 6-5 shows the diagram of the experiment system. Two Vivaldi 004 antennas are
placed at both side of the sand flow with the distance dobj = 17.5 cm. The distance between antennas and the bottom of
the funnel h = 7 cm.

6.2.2 Result and Discussion

The result of the measurement is shown in Figure 6-6. The blue and the red lines are the mean value of all frames that
collected without and with sand flow respectivley. It can be seen that, unlike in the previous expirement, two curves
are almost overlapped and it is hardly to see any difference.

45
 Figure 6-6 Measurement results for transmission detection of the sand flow.

This result can be explained by the difference between the density of the sand flow in the experiment and the sand
flow generated by the sand box. In the experiment, the funnel has a 1 cm diameter hole at the bottom. The sand flow
drops from the funnel has much higher density than the one blow out by the sandbox. Therefore, both reflection and
transmission can hardly show any significant difference. In conclusion, the Novelda UWB radar system might not be
able to provide a satisfied solution to this problem.

46
47
7 Summary and Future Work
Through this thesis work, several UWB radar applications are investigated and the corresponding signal processing
algorithms are developed with good results, and one application (sand flow detection) is still open to good solutions.
The summary is as follows:

• The breath detection has very stable performance that can work right in front of a human, and has a potential to
work through a wall.
• The radar system is able to detect small movements, such as the movement in the experiment with ±0.5mm. The
wavelet packets and EMD methods are shown to be very helpful in heart rate detection with suppressing breath
interference, though the detection is sensitive to irregularly body movement.
• The radar system has been shown to have a high resolution in distance. It can track a quick moving object in the
close range that have a relatively strong reflection.
• The radar system can detect a substance that has a strong reflection or can absorb the incident wave significantly.
• We have not yet found a good solution to detect the low density sand flow that generated by the sand box, due to
the low reflection and high transmission.

A future work could be done with two or three radar modules at the same time, which could enable a two dimensions
or even three dimensions resolution. Therefore, by using the correlation between the signals from different radar
modules, a higher signal to noise ratio could be obtained. Besides that, more application might be investigated, such
as multi-dimensions positioning, tracking, target recognition.

48
49
8 References
[1] Ultra-wideband Transmission Systems, Federal Communications Commission, federal register, title 47, vol.1,
2002.
[2] Ian Oppermann, Matti Hamalainet, Jari Linatti, UWB Theory and Applications, Wiley.com, 2006.
[3] Ben Allen, Mischa Dohler, Ernest Okon, Wasim Malik, Anthony Brown, David Edwards, Ultra Wideband
Antennas and Propagation for Communications, Radar & Imaging, Wiley.com, 2006.
[4] Yinan Yu, Jian Yang, Tomas McKelvey and Borys Stoew, “A Compact UWB Indoor and Through-Wall
Radar with Precise Ranging and Tracking”, International Journal on Antenna and Propagations, vol. 2012,
Article ID 678590, 11 pages, 2012.
[5] J. Yang and A. Kishk, “A novel low-profile compact directional ultra-wideband antenna: the self-grounded
Bow-Tie antenna”, IEEE Trans. on Antennas Propag., vol. 60, no. 3, pp. 1214-1220, March 2012.
[6] J. Yang and A. Kishk, “The self-grounded Bow-Tie antenna”, 2011 IEEE international Symp. on Antennas
Propag., Spokane, Washington, USA, July 3-8, 2011.
[7] S.Fayazi, H.-S. Lui and J. Yang, “Microwave Imaging of Near-Field Object Using Ultra-Wideband Synthetic
Aperture Radar Algorithm”, 2012 IEEE international Symp. on Antennas Propag., Chicago, IL, USA, July 8-
14, 2012.
[8] A-Rawi, Jian Yang, Charlie Orlenius and Magnus Franzen, “The Double-sided 4-port Bow-tie antenna: A
new compact wideband MIMO antenna”, 7th Eur. Conf. on Antennas Propagat. (EuCAP2013), Gothenburg,
8-12 April 2013.
[9] Fayazi, Jian Yang, Hoi-Shun Lui, “UWB SAR imaging of near-field object for industrial process
applications”, 7th Eur. Conf. on Antennas Propagat. (EuCAP2013), Gothenburg, 8-12 April 2013.
[10] Shirin Abtahi, Jian Yang, and Stefan Kidborg, “A new compact multiband antenna for stroke diagnosis
system over 0.5-3 GHz”, Microwave and Optical Technology Lett., Vol. 54, No. 10, pp. 2342–2346, October
2012.
[11] Vivaldi 006 Anti-podal directional antenna, Imego AB
[12] Vivaldi 004 Directional antenna, Imego AB
[13] Gabriel Rilling, Patrick Flandrin, Paulo Goncalves, “On Empirical Mode Decomposition and its Algorithms”,
IEEE-EURASIP Workshop on Nonlinear Signal and Image Processing NSIP, vol. 3, pp. 8-11, 2003.
[14] Donald B. Percival and Andrew T. Walden, Wavelet Methods for Time Series Analysis, vol. 4, Cambridge
University Press, 2006.
[15] http://www.novelda.no, Novelda AS
[16] NVA6100 Preliminary Datasheet, Impulse Radar Transceiver System, Novelda AS
[17] N.E. Huang, Z. Shen, S.R. Long, M.L. Wu,H.H. Shih, Q. Zheng, N.C. Yen, C.C. Tung and H.H. Liu, “The
empirical mode decomposition and Hilbert spectrum for nonlinear and non-stationary time series analysis,”
Proc. Roy. Soc. London A, Vol. 454, pp. 903–995, 1998

50
51
9 Appendix A: Demo system
In order to demonstrate the possible applications the radar system could be used in, a demo system has be made. The
system is based on the APIs offered by Novelda for the radar system. The GUI and signal processing algorithms are
written with Matlab.

9.1 Raw Data

Show the unprocessed raw data from the radar system in the frame view.

Figure 9-1 Screen shot of raw signal from radar.

52
9.2 Clutter Removal

Show the clutter removed signal with SVD method in the frame view. [4]

Figure 9-2 Screen shot of a clutter removed signal.

53
9.3 One Dimension Tracking

A function based on the previous work [4], which achieved one dimensional tracking by using Kalman filter.

Figure 9-3 Screen shot during one dimension tracking on a moving metal board in the radar range.

54
9.4 Breath Detection

A function variant from the off line breathes detection method. The function provides the real time breath rate by
following procedure:

• Collect signal until the data matrix reach certain size;

• Start the loop below until the system quit this function:
1. Update the breath rate in every n seconds(where n is adjustable) by following steps:
 Calculate the frame rate and other key value for further processing;
 Detect breath rate;
2. Drop the oldest signal and put the new signal at the end of the data matrix.

Figure 9-4 Screen shot during breath detection on a human in the radar range.

55
Figure 9-5 Screen shot during breath detection without any human in the radar range.

56
9.5 Heart Rate Detection

A function variant from the off line heart rate detection by wavelet packets. The function provides the heart rate in
every n seconds (where n is adjustable) by a loop with following procedure:

• Collect signal for n seconds and save to a data matrix;

• Calculate the frame rate and other key value for further processing;
• Calculate the nodes that are needed for reconstruction;
• Detect heart rate;
• Drop the data matrix.

Figure 9-6 Screen shot during heart rate detection on a human in the radar range.

57
Figure 9-7 Screen shot during heart rate detection without any human in the radar range.

58
10 Appendix B: Paper Published in EuCAP 2014
There is a paper published in EuCAP 2014 regarding this work.

59
60
 Detection of Breathing and Heartbeat by Using a
Simple UWB Radar System
Qiuchi Jian*, Jian Yang*, Yinan Yu*, Peter Björkholm†, Tomas McKelvey*
*
Dept. of Signals and systems, Chalmers University of Technology, SE-41296 Gothenburg, Sweden
[email protected], (jian.yang, yinan, mckelvey)@chalmers.se
†
Acreo Swedish ICT AB, Gothenburg, Sweden, [email protected]

Abstract—We present the development on an ultra-wideband

(UWB) radar system and its signal processing algorithms for
detecting human breathing and heartbeat in the paper. The
UWB radar system consists of two (Tx and Rx) antennas and one
compact CMOS UWB transceiver. Several signal processing
techniques are developed for the application. The system has
been tested by real measurements.

Index Terms—UWB radar application; wavelet; empirical

mode decomposition, breathing monitoring; heartbeat detection.

I. INTRODUCTION
Ultra-wideband (UWB) radars are widely used in different
applications, such as through-wall tracking and detection [1],
medical monitoring instruments [2], life detection [3], industry
processing monitoring [4], etc. Several distinct advantages
make UWB radar systems attractive: i) good time domain
resolution and therefore accurate tracking and positioning; ii)
strengthened target recognition; iii) robust immunity to passive
jamming; iv) relatively low cost.

In this paper, a simple compact UWB radar system [5] is

applied to the application of breathing and heartbeat detection. Figure 1. Experimental setup of breath detection (upper) and the Vivaldi
antenna (lower).
An experimental vibration system with a known frequency is
set up as an emulating model for heartbeat in order to develop
signal processing algorithms. Based on the experiment data and
true human measurements, several signal processing
algorithms, such as wavelet, empirical mode decomposition,
are explored and implemented. Promising results are achieved.

II. SYSTEM CALIBRATION

System calibration is applied to clarify the system property
and identify potential interferences that could compromise the
measurements, such as temperature effect and jamming. The
results conclude that the interferences could have significant
impact on measurements of minor movement such as human
heartbeat. Thus, adjustments corresponding to these have been Figure 2. Signal processing procedure diagram for breath detection.
done on the further experiment environment setups, including
discarding data collected during heat up period and performing In the experiment phase, measurements have been done for
measurements in electromagnetic shielded environment. 60 seconds continuously. Then the collected raw data has been
processed following the procedure in Fig. 2.
III. BREATHING MONITERORING
As depicted in Fig.2, based on the distance between the
The setup for breathing detection is as shown in Fig. 1: two human and the antennas, it is possible to extract signal from the
Vivaldi antennas are placed in parallel with da = 12 cm, and area that covers the human’s position. Then, by applying
one person is sitting in a chair with a distance dobj = 24 cm singular value decomposition method to remove the reflection
between the chest and the outer edge of the antennas. from the static objects in the signal (please refer to [1]), an

This work was sponsored by Acreo Swedish ICT AB, Gothenburg, Sweden.

61
enhanced clutter-removed signal is obtained. The clutter- techniques, including band-pass filtering, empirical mode
removed signal is been further processed in order to do the decomposition (EMD) and wavelet packets, are explored and
movement detection as well as the breath signal reconstruction. applied to different scenarios, as shown in Fig.6.

A differential phase method is applied here for movement

detection: for two identical signals with a time delay, the time
delay could be obtained by Fourier Transform [1].

Suppose s1 and s2 are identical signals with a phase shift:

s2(τ) = s1(t – τ) (1)
By applying Fourier Transform we have:
S1(ω) = S2(ω) e-jωτ (2)
Where τ is the time shift between two consecutive
measurements, which is positive when the patient is exhaling
and negative when inhaling. Therefore:
S1(ω) / S2(ω) = e-jωτ (3)
Thus, the angel of S1(ω) / S2(ω) could be obtained by:
angel = -jωτ (4)
Figure 4. Breath signal reconstruction from measurements during 55
Since human breath is a slow periodic movement, the seconds.
statistical distribution of the reflection from the breathing chest
is centered at a low frequency. Fig. 3 is the comparison of the
movement detection results with and without presence of a
person. The breathing can be detected by a simple Generalized
Likelihood Ratio Test (GLRT).

Figure 3. Signal extracted for movment detection.

Figure 5. Experimental setup of vibrating corner reflector.
Results from one minute continuous measurements can be
found in Fig. 4. After preliminary processing, the breathing and The result of a moving metal plate with 1.5 Hz frequency is
heartbeat signals are assumed to be additive as shown in Fig. 4. shown in Fig. 7. Although the reconstructed signals share high
The ‘envelop’ is considered as the breathing signal and the similarity in frequency domain, the signal processed by the
higher frequency components is the heartbeat. FIR1 filter contains side lobes with higher amplitude.
However, it can be used as a complementary approach in later
IV. HEARTBEAT DETECTION experiments.
First, human heartbeat movements are emulated by using a
vibrating corner reflector, as shown in Fig. 5. The experimental Then, the tested approaches are applied to the detection of
setup contains a function generator, a vibrator, a corner human heartbeats with the same setup as breathing detection
reflector, and our UWB system. To rule out jamming from presented in Fig. 1 (a). The only difference is all measurements
other sources in order to develop and test signal processing are done in MS room like the simulation experiment above in
(SP) algorithms in the early stage of the work, the experimental order to rule out Jamming. An adjusted signal processing flow
setup, except the desktop computer, was placed inside a was used, as shown in Fig.8.
magnetic shielded (MS) room. Different signal processing (SP)

62
  Compute the mean values between the maxima and
minima curves, and make a residual curve r(t) which is
the pink one in the figure.

 Get the detail known as IMF by d(t) = s(t) − m(t).

 Repeat the process above on the residual r(t) until it fits

the stopping criteria.

The stopping criteria can be defined based on the

requirement of the analysis. In this project, an existing library
is used for calculation which defines the calculation stops once
all condition below are satisfied [7]:
Figure 6. Signal processing procedure diagram for corner reflector movment
detection.
 For each point, meanAmplitude(t) < 0.5* amplitude(t);
 mean of boolcriteria(t) less than 0.05;
 There are less than two extremes in the curve.
where
meanAmplitude(t) = abs(maxAmplitude(t) + minAmplitude(t))/2 (5)
amplitude(t) = abs(maxAmplitude(t) - minAmplitude(t))/2 (6)
boolcriteria(t) = (meanAmplitude(t)/amplitude(t)) > 0.05 (7)

Figure 7. The results of signal of one sample with 1.5 Hz movement

processed in the frequency domain.

Figure 9. Four IMFs that decomposed from the clutter removed signal,
which are used for heartbeat signal reconstruction.

Fig. 9 shows four IMFs that are decomposed from the clutter
removed signal. By combining them together, a rough
reconstructed heartbeat signal could be obtained [8].

Another technique is the wavelet decomposition. The

wavelet decomposition is a technique for decomposing time
Figure 8. Signal processing procedure diagram for heartbeat detection.
series with different scales and times. [9] By using Wavelet
As depicted in Fig. 8, two methods are applied to the clutter transform, a signal could be divided into two parts: Signals in
removed signal: EMD and Wavelet Packets. EMD is a lower frequency (refereed as A, which stands for
nonlinear technique that can decompose a signal into finite ‘Approximation coefficients’) and higher frequency (referred
components known as Instantaneous Mode Functions (IMF), as D, which stands for ‘Details’).
which was developed by N.E. Huang in 1998. Specifically, the
algorithm used in this paper achieves the target signal s(t) By using wavelet packets, the wavelet transform is first
applied to the original signal to get A and D. Then the
decomposition by the following steps [6] :
transform is applied again on both A and D and their outcome
 Find all minima and maxima extremes and plot a curve respectively. Hence, this analysis will produce a binary tree
for each set by using cubic spline interpolation, which with root node on the top of the tree and 2n on the nth level of
are the blue and red curves respectively. the tree. By doing this, the frequency components can be
selected according to our interests.

63
 open area or even through a wall due to its penetration ability.
From the experiments using the corner reflector and
measurements of real heartbeats, we conclude that it is possible
to detect small movements with ±0.5mm amplitude using the
radar system. The wavelet packets and EMD methods are
applied to suppress clutters and interference and the heartbeats
are reconstructed by periodic sinusoid.

Although there are many previous studies focusing on

certain vital signs [8] or signal to noise and clutter ratio
improvement [10], but with the performance of current system,
by combining different signal processing methods, the system
can achieve effective measurement both off-line and in real
Figure 10. Three sub signal that decoposed from the clutter removed signal,
which are used for heartbeat signal reconstruction. Node (4, 1) indicates the time with less power and smaller antennas. Therefore, better
node is the 1st node on the 4th level of the binary tree. mobility is achieved.

Fig. 10 shows three nodes that are decomposed from the As future work, more radar modules can be utilized
clutter removed signal. In order to reconstructed heartbeat simultaneously to achieve better resolutions in two or three
signal, we can combine all signals from nodes that are within dimensions for ranging and tracking applications. By using the
our interest frequency range. correlation between the signals from different radar modules, a
higher signal to noise ratio could be obtained.

VI. REFERENCES
[1] Yu, Yinan; Yang, Jian; McKelvey, Tomas; Stoew, Borys, “Compact
UWB indoor and through-wall radar with precise ranging and
tracking,” International Journal on Antenna and Propagations, vol.
2012, Article ID 678590, 11 pages, 2012.
[2] S. Abtahi, J. Yang, and S. Kidborg, “A new compact multiband
antenna for stroke diagnosis system over 0.5–3 GHz,” Microwave and
Optical Technology Lett., vol. 54, no. 10, pp. 2342–2346, October
2012.
[3] B. Allen, M. Dohler, E. Okon, W. Malik, A. Brown, and D. Edwards,
Ultra wideband antennas and propagation for communications, radar
and imaging, Wiley.com, 2006.
[4] S. S. Fayazi, H.-S. Lui, and J. Yang, “Microwave imaging of
nearfieldobject using ultra-wideband synthetic aperture radar
algorithm,” in Proc. 2012 IEEE Antennas and Propagation Society
International Symposium, 2012, pp. 1–2.
[5] Y. Yu, S. Maalik, J. Yang, T. McKelvey, K. Malmstrom, L. Landen,
and B. Stoew, “A new UWB radar system using UWB CMOS chip,”
Figure 11. Heartbeat signal reconstruction from continuous measurements. in Proc. 5th European Conference on Antennas and Propagation
(EUCAP2011), 2011, pp. 771–775.
In the reality, the breathing and the heartbeat signals are [6] N. E. Huang, Z. Shen, S. R. Long, M. L. Wu, H. H. Shih, Q. Zheng,
additive together in the receiving signals, as shown in Fig. 4. N. C. Yen, C. C. Tung and H. H. Liu: “The empirical mode
decomposition and Hilbert spectrum for nonlinear and nonstationary
The 'envelop' is considered as the breathing signal and the time series analysis,” Proc. Roy. Soc. London A, vol. 454, pp. 903–
higher frequency components is the heartbeat which is of our 995, 1998.
interest. By applying EMD and wavelet packets methods, both [7] Gabriel Rilling, Patrick Flandrin, Paulo Goncalves: On Empirical
the breathing and the unwanted components with higher Mode Decomposition and its Algorithms, 1998.
frequencies than the heartbeat are sufficiently suppressed. One [8] E. P. Souza Neto, M. A. Custaud, C. J. Cejka, P. Abry, J. Frutoso, C.
Gharib and P. Flandrin: “Assessment of cardiovascular autonomic
example of the reconstruction results in time domain is shown control by the Empirical Mode Decomposition,” in Proc. 4th Int.
in Fig. 11. We can see that the outcome of both methods share Workshop on Biosignal Interpretation, Como (I), 2002, pp. 123-126.
high similarity on their shapes and synchronized at their peaks. [9] Donald B., Percival & Andre T. Walden: Wavelet Methods for Time
Series Analysis, 2000.
V. CONCLUSIONS [10] A. Nezirovic, A. G. Yarovoy, and L. P. Ligthart, “Signal processing
for improved detection of trapped victims using UWB radar,” IEEE
In this work, possible applications are implemented and Transactions on Geoscience and Remote Sensing, vol. 48, no. 4, pp.
tested using one UWB radar chip. Breath detection is one of 2005-2014.
the most promising possibilities for patient monitoring in an

64
1