Ghost target classification

using scene models in radar

Anton Sedin
David Wadmark

Department of Automatic Control

MSc Thesis
TFRT-6140
ISSN 0280-5316

Department of Automatic Control

Lund University
Box 118
SE-221 00 LUND
Sweden

© 2021 by Anton Sedin & David Wadmark. All rights reserved.

Printed in Sweden by Tryckeriet i E-huset
Lund 2021
Abstract

In surveillance contexts, radars can be used to monitor an area, detecting and track-
ing moving objects inside it. Monitored areas in urban environments often contain
many surfaces that reflect radar waves, which can have the undesired consequence
of a single object producing multiple tracks due to multipath propagation effects.
This thesis considers a method of identifying if a track is produced by a real ob-
ject, or if it stems from multipath effects. The proposed method works by creating
a machine-learning-based classifier and modelling the monitored scene over time.
Tracks are assigned features based on their characteristics and the state of the scene
model in regards to their position. These features are then used as inputs to the
classifier model to produce the classification. We propose four machine-learning-
based classifier models, with two different sets of structures and features used. The
classifier models are compared to a naive classifier model for reference.
The proposed models all outperform the naive classifier, although some of them
are biased. As for the usefulness of the scene model, the results are mixed but show
promise. We believe that the scene model can improve classification performance
further with more and better data.

Keywords
radar surveillance, multipath, machine-learning, classification

3
Acknowledgements

We acknowledge the fantastic support and council that was provided by our in-
dustry supervisors Dr. Sebastian Heunisch and Dr. Aras Papadelis, as well as our
unofficial industry supervisors Dr. Stefan Adalbjörnsson, Dr. Anders Mannesson
and Daniel Ståhl. We would like to thank them for all the great discussions, remarks
and advice, and for providing superb feedback on the report with short notice.

We want to thank our supervisors at Lund University, Prof. Andreas Jakobsson

and Dr. Björn Olofsson for providing great advice and feedback on the report. We
also want to thank our examiners Prof. Rolf Johansson and Senior Lecturer Johan
Lindström for critiquing and grading this thesis.

Additionally, we would like to thank our opponents David Carpenfelt and Gustaf
Broström for their fair critique, and André Nüßlein for proofreading the report.

Finally, we want to give a heartfelt thank you to our families for their uncondi-
tional support.

5
Abbreviations

FMCW - Frequency Modulated Continuous Wave

IF - Intermediate Frequency
ANN - Artificial Neural Network
RCS - Radar Cross Section
SNR - Signal-to-Noise Ratio
FFT - Fast Fourier Transform
DFT - Discrete Fourier Transform
LOGOCV - Leave-One-Group-Out Cross-Validation

7
Contents

1. Introduction 11
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 The multipath problem . . . . . . . . . . . . . . . . . . . . . . 11
1.3 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Previous work . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2. Theory 13
2.1 Radar fundamentals . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 FMCW radar . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Tracker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Multipath fundamentals . . . . . . . . . . . . . . . . . . . . . . 18
2.5 Machine learning methods . . . . . . . . . . . . . . . . . . . . . 19
2.6 Scene mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.7 Evaluation metrics . . . . . . . . . . . . . . . . . . . . . . . . . 24
3. Data 25
3.1 Gathering data . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Processing and annotating data . . . . . . . . . . . . . . . . . . 26
3.3 Structure of dataset . . . . . . . . . . . . . . . . . . . . . . . . 27
4. Scene model 28
4.1 Scene-specific information . . . . . . . . . . . . . . . . . . . . 28
4.2 Mapping reflective surfaces . . . . . . . . . . . . . . . . . . . . 29
4.3 Feature density maps . . . . . . . . . . . . . . . . . . . . . . . 30
5. Features 32
5.1 Track-specific features . . . . . . . . . . . . . . . . . . . . . . . 32
5.2 Scene-specific features . . . . . . . . . . . . . . . . . . . . . . . 33
6. Model structures and evaluation 34
6.1 General structure . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.2 Naive classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.3 Model 1 — Random forest with only track-specific features . . . 35
6.4 Model 2 — Random forest with scene-specific features . . . . . 35

9
Contents

6.5 Model 3 — Multilayer perceptron with only track-specific features 36

6.6 Model 4 — Multilayer perceptron with scene-specific features . . 36
6.7 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.8 Feature selection . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.9 Model parameter tuning . . . . . . . . . . . . . . . . . . . . . . 37
7. Results 39
7.1 Classification results . . . . . . . . . . . . . . . . . . . . . . . . 39
7.2 Visualization of scene models . . . . . . . . . . . . . . . . . . . 42
8. Discussion 45
8.1 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
8.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
9. Conclusion 48
Bibliography 49

10
1
Introduction

1.1 Background
Short-range radar sensors have become a field of interest in the surveillance industry
in recent years. Unlike video surveillance systems, radar-based surveillance systems
can operate independent of lighting conditions and they do not suffer from the same
large decrease in performance in foggy, rainy and snowy conditions. Objects are
detected by transmitting an electromagnetic signal from the radar, and collecting
the reflections from the surroundings. By using certain transmit patterns and com-
bining multiple receiver antennas in an array, the range, radial velocity and angle of
detected objects can be deduced. However, there is no direct way to tell if a signal
comes from a direct reflection from the object or if the signal has been reflected on
other surfaces. This is a big problem in radar surveillance systems that are mounted
in urban environments with many reflecting surfaces like walls, containers or cars.

1.2 The multipath problem

The problems caused by ambiguous reflections of radar signals can be illustrated
by Figure 1.1. In this fictive scenario, the radar is able to trace the target over time
by measuring the position and the velocity of the target. We will refer to this trace
as a track. Radar signals can also be reflected in other surfaces like the wall in this
case, either before or after hitting the target. We will refer to this as the multipath
phenomenon. This can cause additional tracks to appear although there is only one
target in view of the radar. These ghost tracks can cause false alarms, for instance
if they appear in a zone where no people are allowed. In any real installation, the
environment is typically more complex than in Figure 1.1 where there is only one
perfectly flat wall. This means that radar signals can often travel in an exceedingly
large number of ways from the radar to a target and back. Therefore, it is not an
easy task to construct an accurate geometrical model of the reflections.

11
Chapter 1. Introduction

Figure 1.1 Illustration of the multipath phenomenon. The radar signals travel along
two different paths back to the radar after reflecting of the target causing two tracks
to appear from one target.

1.3 Purpose
This thesis aims to investigate the possibility of using radar data to reliably clas-
sify whether or not a detected target is in fact real, or a result of the multipath
phenomenon. More specifically, we will try to solve this problem using machine-
learning methods, and investigate whether the classification can be aided by con-
structing a model of the monitored scene over time.

Scope
We make no attempt to distinguish between the type of objects such as pedestrians,
cars and bikes in this thesis, only whether the detected target is a ghost target or a
real target. All objects are treated the same, no matter what type of object it is. We
also assume that the radar is stationary. The investigation will be done using a single
radar sensor with proprietary signal processing and tracking software provided by
our industry supervisors. This thesis will not consider any changes to these systems.

1.4 Previous work

Previous attempts have been made to classify ghost targets from radar signals, but
predominantly in automotive settings where the radar is mounted on a vehicle, and
is therefore not stationary [Ryu et al., 2018], [Prophet et al., 2019], [Kraus et al.,
n.d.] Attempts have also been made to infer the position of walls using a stationary
radar by making some assumptions about multipath signals [Nüßlein, 2021].

12
2
Theory

An introduction to radar signal processing and some machine-learning concepts will

be presented in this chapter for readers unfamiliar with these concepts. The first sec-
tions will cover the fundamentals of radar technology and radar signal processing.
Then, the tracker used in this thesis and the multipath problem are described in more
detail. Finally, some concepts from machine learning that are used in this thesis are
presented.

2.1 Radar fundamentals

Radar is an acronym for Radio Detection and Ranging. The technology was orig-
inally developed to detect and locate aircraft in World War II [Sim, 2014] . Nowa-
days, radar technology is used in a wide range of applications such as flight control
systems, radar astronomy and meteorological precipitation monitoring.
Radar systems use radio waves to determine the range, angle or velocity of
objects. They are composed of a transmitter and receiver. The transmitter emits
radio waves which in turn are reflected or scattered by objects, and the receiver
captures the reflected or scattered waves. There are many types of radars which use
different transmitting schemes, but in this thesis exclusively Frequency Modulated
Continuous Wave (FMCW) radars were used, which will be covered in the next
section.

2.2 FMCW radar

Range measurement
To measure the range from the radar to an object, an FMCW radar transmits a signal
called a chirp [Smith, 1997]. A chirp is a sinusoid where the frequency increases
linearly with time. The frequency and amplitude of a chirp as a function of time can
be seen in Figure 2.1.

13
Chapter 2. Theory

f Amplitude

t t

Figure 2.1 Illustration of a chirp. Left: frequency as a function of time. Right:

amplitude as a function of time.

f Tx Rx Tx Rx
B

∆f

τ tc t

Figure 2.2 Transmitted signal (Tx) and received signal (Rx).

If there is an object within range of the radar, the chirp will be reflected from the
object and a time-delayed chirp will be received by the radar. This transmit-receive
pattern is illustrated in Figure 2.2.
The received signal is then mixed with the transmitted signal to a so-called inter-
mediate frequency (IF) signal. This is a sinusoid where the frequency and phase is
the difference of the frequencies and phases of the received and transmitted signal:

xIF (t) = sin [(ω2 − ω1 )t + (φ2 − φ1 )] (2.1)

where ω1 and ω2 are the frequencies of the transmitted and received signal, and φ1
and φ2 are the phases of the transmitted and received signal.

14
 2.2 FMCW radar

The IF signal will have frequency

2Sd
f0 = ω2 − ω1 = Sτ = (2.2)
c
which is equivalent to ∆ f in Figure 2.2. Here, S is the rate of change of the chirp
frequency, d is the distance to the interfering object, and c is the speed of light. From
Equation (2.2) the distance from the radar to an interfering object can be deduced.
We will refer to this distance as the range of an object.

Doppler
To deduce the radial velocity of an object, multiple chirps are transmitted in rapid
succession. With the assumption that the object is traveling much slower than the
speed of light, the resulting IF signals of each chirp will have approximately the
same frequencies, but their phases (φ2 − φ1 in Equation (2.1)) will be different for
non-zero radial velocities. This is because the distance d to the object changes be-
tween chirps due to the movement of the object. This change in distance ∆d is small,
but significant in relation to the wavelength λ of the chirp. The relation between
phase difference ∆φ and difference in distance ∆d is [Rao, 2017]

4π∆d
∆φ = . (2.3)
λ
By transmitting chirps separated by a time of tc and inserting the relation ∆d =
vtc in Equation (2.3) we have

4πvtc λ ∆φ
∆φ = =⇒ v = (2.4)
λ 4πtc

where v is the radial velocity of an object. The signals received by transmitting a set
of chirps will be referred to as a frame from here on.

Azimuth
To determine the angle of arrival θ of an object, several equally spaced receiver
antennas are used. The incoming signals are assumed to be plane waves. The sce-
nario is illustrated in Figure 2.3, where the incoming signals to each of the receiver
antennas (black rectangles) are represented by the red arrows.
For non-zero angles, the received signals at each of the antennas will have a
difference in phase which we call ω. From this phase difference, the angle of arrival
of the received signal can be deduced using the relation
 
2πd sin θ λω
ω= =⇒ θ = arcsin . (2.5)
λ 2πd

15
Chapter 2. Theory

θ θ d sin θ

Figure 2.3 Incoming signal (red arrows) from an object at an angle θ to some
receiver antennas (black rectangles).

Fourier transforms
When several targets appear in front of the radar the transmitted chirps are reflected
multiple times, leading to the IF signal becoming a sum of sinusoids. To identify
the individual sinusoids, a Fourier transform is applied to convert this time-domain
signal into the frequency domain. From here the independent sinusoids that make
up the IF signal can be deduced by looking at where the peaks of the transformed
signals are. Since the converted signal is complex, each value containing both an
amplitude and a phase, we can also get the initial phase of each sinusoid by looking
at the phase of its peak in the frequency domain.
Doing this in practice means using a fast Fourier transform (FFT) [Duhamel
and Vetterli, 1990], which is a computationally efficient algorithm that computes
the discrete Fourier transform (DFT) of a sequence. When an IF signal is generated
from a receiving antenna, it is transformed using an FFT and the results are stored
in a vector. This is repeated for each chirp in a frame, the resulting vectors being
stored as rows in a matrix. For each range bin another FFT is applied, creating a
range-velocity plot. These steps are repeated for every receiver antenna, creating a
3D-array where a final FFT can be applied for each range-velocity bin. Thus, the
range, radial velocity, and azimuth angle of individual objects can be detected in a
scene with multiple reflections. These steps can be seen in Figure 2.4. For further
information, we refer to [Rao, 2017].

Signal strength
There are two measures commonly used in radar sensors for describing the strength
of the reflected signal. The radar cross-section (RCS) denoted by σ , is a measure-
ment of how well an object reflects radio waves. It interprets the object as if it were a
perfectly reflecting sphere, the cross-sectional area of which is the dimension of the
measure. While RCS formally describes the detected object rather than the strength
of the recieved signal, the two are correlated as can be shown by the radar range
equation [Richards, 2005]:

Pt Gt σ Ae (4π)2 r4 Pr
Pr = 2 4
=⇒ σ = (2.6)
(4π) r Pt Gt Ae

16
 2.3 Tracker

Figure 2.4 Signal processing steps. (top left) range-chirp index matrix, (top right)
range-velocity matrix, (bottom left) range-velocity-antenna cube, and (bottom right)
range-velocity-azimuth cube.

where Pt and Pr are, respectively, the power to the transmitting antenna and the
received power. Ae is the effective area of the receiving antenna, a measure of its
geometric area and efficiency. Finally, Gt is the gain of the transmitting antenna,
and r is the distance from the radar to the object.
Another, more straight-forward way to describe the strength of the reflected
signal is the signal-to-noise ratio (SNR). For a bin i, in a set of bins I, we define this
ratio as:
Pr,i
SNRi = (2.7)
P̃r,I
where Pr,i is the received power in a specific bin and P̃r,I is the mean noise power
estimated here as the median of the received power in all bins. Note that SNR does
not take the distance to the detected object into account, meaning a small object
close to the radar could have a higher SNR than a large object further away.

2.3 Tracker
The purpose of the tracker is to follow the position of a target over multiple frames.
The high velocity resolution of millimeter wave radars means that an object like a
car, person, or even a small animal will give off multiple detections. This is not only

17
Chapter 2. Theory

due to the spatial extent of the object, but also due to different parts of the detected
object having different radial velocities, for example the swaying limbs of a person.
This phenomenon is known as a micro-Doppler effect [Chen et al., 2006]. After the
Fourier transforms, each frame contains a point cloud of detections, with each point
containing polar coordinates, radial velocity, RCS and SNR values. By grouping
these detections into clusters and connecting them over time, tracks are created. In
this thesis, this was done by proprietary tracking software, which we were given
access to by our industry supervisors. We will refer to this software as the tracker.

Clustering and association

The first step in the tracking process is to determine which detection points stem
from the same object, and group them together. This is called clustering and is done
using a process that takes the position and RCS values of the detections into ac-
count. Next these clusters are given possible associations to existing tracks in the
previous frame based on their velocities and positions. Finally, out of these possi-
ble associations, the best combination of tracks and clusters are chosen based on
minimizing the distances between them.

Updating tracks
Updating the existing tracks is based on a version of the well-known Kalman fil-
ter [Kalman, 1960]. If a track has an associated cluster, the new position will be
a weighted average of its predicted position and the position of the cluster. In
cases where a track has no association to a cluster, the tracker will perform dead-
reckoning to predict the current position of the object. If the track is not associated
with a new cluster frequently over a certain amount of frames, it will no longer be
considered alive, and subsequently be removed. Any remaining clusters that have
not been associated to any existing track are used to create new tracks.

2.4 Multipath fundamentals

Given a radar sensor with an object visible to it, the direct path is the straight line
from the radar to the object. The multipath phenomenon is the result of the radar
signal not traveling along this direct path, instead taking an indirect path after being
reflected by one or more other surfaces in the scene. The signal can take an indirect
path before and/or after being reflected off the object, which will alter the position
of the ghost object. Since the perceived range to the ghost target is equal to the
distance the signal has traveled, and the direct path is the shortest, a ghost object
will always be further away than the real object that created it. Therefore, given a
frame with multiple tracks, it can be assumed that the one closest to the radar is a
real track. For a more detailed description of the geometrical aspect of the multipath
phenomenon we refer to [Nüßlein, 2021].

18
 2.5 Machine learning methods

As an electromagnetic signal always loses power from a reflection on most ma-

terials, there is a limited number of reflections that can occur before the signal can
no longer be distinguished from noise by the radar. It also means that the RCS of
a real track will always be larger than the RCS of any ghost tracks that have been
generated from that same object. Because of this it can also be assumed that the
track with the largest RCS value in a scene is a real track. It is worth noting that the
assumptions outlined in this section are not always true, as an object which is not
visible by the radar can produce reflections in surfaces that are visible to the radar.

2.5 Machine learning methods

Decision tree classifiers
Given a set of samples described by feature vectors x and a set of labels y which de-
scribes the categorical value of the feature vectors, a decision tree classifier [James
et al., 2017] can be used to predict the label y for a feature vector x. When fitting
a decision tree to a dataset (a set of feature vectors and corresponding labels), the
decision tree recursively splits the dataset into two sets by examining the value of
some feature, where the split minimizes or maximizes some metric. In this thesis,
the metric used was the Gini impurity [Laber and Murtinho, 2019]. With a set of
samples having J different possible labels, and letting pi denote the proportion of
samples labeled with class i, the Gini impurity for this set is calculated as
!
J
IG = ∑ pi ∑ pk . (2.8)
i=1 k6=i

In a binary classification problem such as the one in this thesis, the Gini impurity
can be simplified to

IG = p1 p2 + p2 p1 = 2p1 p2 = 2p1 (1 − p1 ) . (2.9)

An optimal split which minimizes the Gini impurity is then a split where one set
only contains samples of class 1 and the other set only contains samples of class
2. An example of a decision tree fit to the dataset used in this thesis can be seen in
Figure 2.5. Two features, "track lifetime" and "closest" which are explained further
in Table 5.1 were used to create the tree, and the depth was limited to 2.

Random forest classifiers

A random forest is an ensemble classifier which combines the result of several sub-
classifiers by majority voting, where the most commonly predicted label by the sub-
classifiers is chosen as the final prediction [James et al., 2017]. The sub-classifiers
in a random forest are decision trees, where in the case of this thesis, the correlation
between the sub-classifiers was reduced by using bagging (Bootstrap Aggregating)

19
Chapter 2. Theory

Figure 2.5 Example of a very simple decision tree fit to the dataset used in this
thesis. The tree has a depth of 2 and only 2 features were used to create the tree.

and boosting (or feature bagging) [James et al., 2017]. Bagging is the practice of
sampling, with replacement, a subset of the dataset used for training, and then us-
ing that subset to fit the model instead of the original training set. Boosting is the
practice of restricting the features considered when making a split in a decision tree
to a random subset of the features. In this thesis, the number of features consid-
√
ered in the random forests used were n f rounded to the nearest integer, where
n f is the number of features in the feature vectors x. This is a common choice for
classification tasks [Hastie et al., 2008].

Artificial neural networks

An artificial neuron is a function which maps an input vector x to an output scalar y
according to
!
m
y(x) = φ (wT x) = φ ∑ w jx j (2.10)
j=0

where w is a vector of weights with dimension m × 1 and φ is an activation function

(see next section). Mapping an input vector x to several neurons will produce an

20
 2.5 Machine learning methods

Figure 2.6 Schematic of a simple artificial neural network.

output vector y with an element yk for every neuron:

!
m
yk (x) = φ (wTk x) = φ ∑ wk j x j . (2.11)
j=0

The result of combining several artificial neurons like this is commonly known as
an artificial neural network (ANN) [Gurney, 1997]. An ANN is usually divided into
layers where the output vector y from one layer is used as input to the next. The
structure of an example ANN is visualized in Figure 2.6. This particular network
accepts 5 input values and produces one output value.

Activation functions
The main purpose of using activation functions in artificial neural networks is to add
non-linearity to the network. Without activation functions, artificial neural networks
are nothing more than linear transformations of the input. Many activation functions
also keep the outputs of neurons from growing in an unbounded fashion. This can
cause ANNs to become unstable and can cause computational problems.
Three main activation functions were used in this thesis; ReLU, sigmoid and
softmax [Sharma, 2017]. ReLU is short for Rectified Linear Unit and is defined as

x if x > 0
ReLU(x) = (2.12)
0 otherwise .
In reality, sigmoid functions refer to a class of functions but in the context of ma-
chine learning it typically refers to the logistic function:
1
S(x) = . (2.13)
1 + exp(−x)

21
Chapter 2. Theory

The softmax function is a generalization of the logistic function beyond the one-
dimensional case. Consider a vector x consisting of K real numbers. The softmax
function σ produces a K-dimensional output where the output at index i is defined
as:
exp(xi )
σ (xi ) = K . (2.14)
∑ j exp(x j )

Supervised learning
With several input vectors xk and corresponding targets yk it is possible to adjust
the weights in an ANN to estimate a function which maps xk to yk . This process is
called supervised learning. For this process two things are necessary; a loss func-
tion and an optimization algorithm. The loss function computes a measure of how
poor the network output ŷk = f (xk ) is in relation to the label yk . A commonly used
loss function in classification contexts is the cross entropy loss function [Brownlee,
2020].
Consider a classification task where c is the target class index, and ŷ is the output
from the network, where each element ŷc in ŷ represents the score of class c. This
score is first approximated as a probability with the softmax function from Equation
(2.14), and then the negative logarithm of the likelihood of the data is calculated:
 
exp(ŷc )
L(ŷ, c) = −wc log . (2.15)
∑ j exp(ŷ j )

Here, wc represents the weight given to class c.

The backpropagation algorithm described in more detail in [Goodfellow et al.,
2016] can be used to calculate the gradients of the loss function with respect to
the weight of every neuron by recursively applying the chain rule of derivatives.
When the gradients are known, the weights of the network can be iteratively up-
dated towards a minimum of the loss function. In this thesis we exclusively used
the network optimization algorithm known as Adam, which is described further in
[Kingma and Ba, 2017].

Dropout
To avoid overfitting the neural networks to the training sets, a regularization tech-
nique known as dropout [Srivastava et al., 2014] was used. In the implementation
of dropout that was used in this thesis, random elements of the input vector x are set
to zero after multiplication with the weights of each layer except for the final layer.
Each element is set to zero according to a Bernoulli distribution Zik ∈ B(p) where
k is the index of the layer and i is the index of the element in the vector. All Zik are
independent. The value of p used in this thesis was p = 0.3.

22
 2.6 Scene mapping

2.6 Scene mapping

In order to build a model of the scene visible to the radar, a map needs to be created.
[Thrun et al., 2006] defines a map m as a list of objects in a scene, along with their
properties:
m = {m1 , m2 , ..., mN } (2.16)
With N objects in the scene, each mn with 1 ≤ n ≤ N specifies a property. Depending
on the type of map, the index n either specifies a certain feature, or a specific location
in the scene.

Occupancy grid maps

Occupancy grid maps are a probabilistic approach to mapping that is popular in
the field of robotics [Siciliano and Khatib, 2007]. The scene is simplified into a
discrete, two-dimensional grid of independent cells, where mi denotes the grid cell
with index i. We define the map as the union of all grid cells:
[
m= mi (2.17)
i

and each cell can be given its own posterior probability of being occupied by an
object:
p(mi |z1:t , x1:t ) = p(mi |z1:t ) (2.18)
where z1:t is the set of all sensor measurements, and x1:t is the path of the robot
[Thrun et al., 2006]. The equality in Equation (2.18) follows from the fact that the
radar is static and its position is considered known in our case. Expanding on this
expression for the occupancy probability we get

p(zt |mi )p(mi |z1:t−1 )

p(mi |z1:t ) = (2.19)
p(zt |z1:t−1 )

and applying Bayes rule to the measurement model p(zt |mi ) we get

p(mi |zt )p(zt )p(mi |z1:t−1 )

p(mi |z1:t ) = . (2.20)
p(mi )p(zt |z1:t−1 )

By calculating the odds of occupancy instead of the probability of occupancy we

can cancel probabilities that are difficult to calculate like p(zt ) and p(zt |z1:t−1 ):

p(mi |z1:t ) p(mi |zt ) p(mi |z1:t−1 ) p(¬mi )

=
p(¬mi |z1:t ) p(¬mi |zt ) p(¬mi |z1:t−1 ) p(mi )
(2.21)
p(mi |zt ) p(mi |z1:t−1 ) 1 − p(mi )
= .
1 − p(mi |zt ) 1 − p(mi |z1:t−1 ) p(mi )

23
Chapter 2. Theory

By taking the logarithm of the odds of occupancy from Equation (2.21) and denoting
it lt (mi ) we arrive at an additive and thus more stable representation of occupancy
in grid cell mi at time t:
p(mi |zt ) p(mi |z1:t−1 ) p(mi )
lt (mi ) = log + log − log
1 − p(mi |zt ) 1 − p(mi |z1:t−1 ) 1 − p(mi )
(2.22)
p(mi |zt )
= log + lt−1 (mi ) − l0 (mi ) .
1 − p(mi |zt )
The expression lt−1 (mi ) represents the log-odds of occupancy of grid cell mi at time
t − 1 and l0 represent the prior log-odds of occupancy of the same grid cell. Thus,
all we need to construct the map is the prior occupancy probability and a way to
calculate p(mi |zt ).
This can be done with a so-called inverse sensor model [Thrun et al., 2006].
Given a measurement, this sensor model updates the map according to a probability
distribution based on the accuracy and characteristics of the sensor. One way to do
this is by raising the occupancy odds in the cells around the measurement point, and
also lowering the occupancy odds in the cells between the measurement point and
the position of the sensor. It is worth noting that the inverse sensor model assumes
that there is no correlation between the occupancy of a cell and the occupancy of its
neighboring cells.

2.7 Evaluation metrics

Classification
To evaluate the classification results, precision, recall and F1 score were used. These
metrics are calculated for each class in the classification problem, and can be av-
eraged over all classes to get a combined evaluation for all classes. Denoting TP
as true positives, FP as false positives and FN as false negatives, we can state the
formulas for calculating precision P and recall R:
TP
P= (2.23)
TP + FP
TP
R= . (2.24)
TP + FN
Precision can be seen as a measure of how many of the selected items that are
relevant, and recall as a measure of how many of the relevant items that are selected.
F1 -score (or traditional F-measure) is the harmonic mean of precision and recall:
P·R
F1 = 2 · . (2.25)
P+R
This measure computes an evaluation of a classification result as a single number,
which is practical. However, it weighs precision and recall equally irregardless of
their consequences for the specific problem and should thus be used with caution.

24
3
Data

A common sentiment in the area of machine learning is that a model can only be as
good as the data it is trained with. This is of course true, as the model can only be
trained to make predictions about samples that are already in the dataset used and
there are no guarantees for the quality of predictions on data from new scenarios and
contexts. Another aspect is that the dataset needs to contain enough samples for the
model to be able to accurately capture the information contained within the current
dataset. Therefore, a large and diverse dataset is essential to produce a robust and
accurate machine learning model.

3.1 Gathering data

Recording Data
The data used in this thesis was recorded with a 60 GHz radar designed to be used
for surveillance. The radar parameters can be found in Table 3.1. The data was col-
lected by recording scenes with a camera and collecting radar data at the same time,
the recorded video being used to aid with annotation. We will refer to a specific
recording from one of these scenes as a scene or a sequence. Some of the scenes re-
semble typical radar installations that a customer might use, but most of the scenes
are chosen specifically to produce at least some amount of multipath signals. The

Table 3.1 Parameters for the radar used to record the data used in this thesis.
Radar type FMCW
Operating frequency 60 GHz
Frame rate 10 frames/s
Maximum range 138 m
Range resolution 0.765 m
Velocity resolution 0.098 m/s
Azimuth accuracy 1◦

25
Chapter 3. Data

recordings can be divided into two subsets; 13 short ones around one to three min-
utes each, and three long ones around 10 minutes each. They mostly feature people
walking in the scene. Some recordings also contain cars and bikes, but in this thesis
we make no attempt to distinguish between the type of objects.

Previously recorded data

In order to minimize the amount of time that had to be spent on recording, we were
given access to a large collection of proprietary data by our industry supervisors.
Unfortunately, only a few scenes were found that were deemed useful for our pur-
poses. The scenes were usually very short, and containing few multipath tracks in
them. Many of them also had their static detections filtered out, which we needed
as will be shown in the next chapter. All the publicly available datasets found were
of no use to us, as they were mostly intended for use in development of autonomous
vehicles, meaning the radar was not stationary.

3.2 Processing and annotating data

All detections above a certain velocity threshold were removed in order to remove
false detections which most likely originated from signal noise caused by hardware
in the radar. This threshold was chosen large enough to include walking persons
and slow-moving cars. The data from each scene was processed by the proprietary
tracking software outlined in Section 2.3 in order to create tracks. These were then
manually annotated using the video recorded from each scene. The tracks which
were derived from a person walking were annotated as real. Tracks which obvi-
ously originated from sources other than real detections or multipath effects were
annotated to be ignored. These were usually generated by stationary objects sway-
ing in the wind, like trees or flags. In scenes where this was prevalent, the areas
where these tracks originated were marked, and all tracks created in those areas
were automatically annotated to be ignored. Finally, all tracks which were not an-
notated as real tracks or flagged to be ignored were automatically annotated as ghost
tracks.
An issue with annotating this way is that it is prone to human error. The camera
used to record the video could not capture the entire scene, meaning people were
sometimes moving outside of its field of view. Even when people are captured by
the camera, it is not always evident which track stems from a real target, as can be
seen in Figure 3.1. Another issue is that tracks can blend together, meaning a track
can start from a multipath effect and then at one point it transitions to tracking a
real target, or vice versa. In this case we chose to annotate the track as real or ghost
depending on what the source of the track was for the majority of its lifetime.

26
 3.3 Structure of dataset

Figure 3.1 Plots of tracks generated by the tracker. A simple frame (left) where
track 2203 is real, and a more problematic frame (right) where it is difficult to tell
which tracks are real.

Table 3.2 The distribution of samples in the dataset.

Samples from real targets Samples from ghost targets Total samples
21 268 72 164 93 432

3.3 Structure of dataset

The dataset used in this thesis is divided into samples where each sample is made up
of features belonging to one track in one frame. These datapoints can for example
be the lifetime of the track, the position of the track and an estimate of its RCS
value. A single track will therefore produce multiple samples, one for each frame
that it is active. It is important to note that classification is done by sample, and not
by track, meaning that a track can be classified as real in one frame and a ghost in
the next frame. The distribution of the samples in the dataset is shown in Table 3.2.

27
4
Scene model

Multipath propagation is dependent on the specific geometry of reflective surfaces

in the scene that is being monitored. Given the simple scenario seen in Figure 1.1,
only ghost tracks can appear beyond the wall to the right of the radar. However, the
radar has no way of knowing this. It could be mounted in a completely different
environment where there is no wall on its right side. In fact there are very few as-
sumptions that can be made about the monitored scene. Therefore, trying to classify
a track using its position without any knowledge of the scene would be ineffective.
Inversely, if certain information specific to a scene can be mapped accurately by the
radar, that information would be a strong tool to use for classification.

4.1 Scene-specific information

One type of scene-specific information that would be of use is the position of large
reflective surfaces, like the walls of a building. If a track appears behind one of
these surfaces it is very likely caused by multipath effects. In most cases these ob-
jects will not change over time, however one could easily imagine a few examples
where this is not the case. If a radar unit was set up to monitor a construction site or
a parking lot for large vehicles, the reflective surfaces could change over time if for
example a large truck parks in front of the radar. Therefore, the scene model imple-
mentation should preferably be able to handle environments that behave somewhat
dynamically.
Another way to utilize information specific to the monitored scene is to map
the areas of the scene where real tracks tend to appear, and where they do not. An
obvious problem here is that we do not know which tracks are real, but we can
create features out of the the assumptions that were derived in Section 2.4 to make
good guesses. In each frame that contains tracks, it is possible to identify which
track is the strongest one, closest one, or has existed the longest. If the position of
these tracks are stored, an approximate map over where assumed real tracks and
ghost tracks frequently appear can be generated.

28
 4.2 Mapping reflective surfaces

4.2 Mapping reflective surfaces

After a literature study, occupancy grid mapping was chosen as the method of mod-
eling the reflective surfaces in the scene. It has an inherent advantage in that it not
only maps areas of the scene which are occupied, but also areas that are unoccupied.
This is of interest as real tracks should only appear in the unoccupied areas of the
scene. Occupancy grid maps are also suited to map a dynamic environment over
time, as shown by their popularity in the field of mobile robotics.
The grid cells in the map were defined in a Cartesian coordinate system where
each grid cell represented an area of 1 m2 . We found that this relatively low resolu-
tion produced the most optimal map from a visual standpoint, while maintaining a
relatively low level of computational complexity.

Extracting static detections

In order to map the reflective surfaces, all detections with an absolute radial veloc-
ity over a certain threshold were filtered and removed. This velocity threshold was
chosen so that only the detections between the smallest positive and the smallest
negative velocity bins were extracted. Since even moving objects can result in de-
tections with no radial velocity, either due to micro-Doppler effects or by moving
concentrically around the radar, a simple filter was implemented which removed any
detections within 2 meters of a track. The remaining detections were then assumed
to be caused by static objects and were used to build a map of the scene.

Occupancy grid map implementation

The gathered data was mapped by our own implementation of an occupancy grid
map, that functioned in the following way. A grid map is created with a given size
and resolution, where each cell has the occupancy probability p0 . For each frame
in the recorded scene, and each static detection not close to a track in that frame,
the grid map is updated with the detection according to the inverse sensor model
described in the next section.

Inverse sensor model

As described in [Thrun et al., 2006], inverse sensor models can be learned by for
example sampling from a sensor model and fitting a neural network to the gener-
ated samples. However, creating the best possible inverse sensor model was not the
focus for this thesis. Instead, we constructed our own inverse sensor model with the
purpose of being simple and computationally efficient. The algorithm is described
below:

1. Receive detection at polar coordinates (r0 , θ0 ).

2. Sample Ns pairs of coordinates from the distributions r ∈ N(r0 , σr2 ) and θ ∈

N(θ0 , σθ2 ).

29
Chapter 4. Scene model

3. For each sampled point, update the grid cell containing the point with proba-
bility p = p0 + αm /Ns according to Equation (2.22).
4. For each sampled point, use Bresenhams line algorithm [Bresenham, 1965]
to find all grid cells in a straight line from the point to the radar at the point of
origin. For all these grid cells, update the occupancy value with probability
p = p0 − βm /Ns according to Equation (2.22).
The parameter p0 is the prior probability of occupancy and σr2 and σθ2 are known
beforehand from the forward sensor model. The user parameters Ns , αm and βm are
chosen with regards to the computational complexity and to how much the occu-
pancy probability is updated when a detection is received. Because the radar in our
model is stationary and detections are accumulated at a high frequency in some grid
cells, αm and βm were chosen to be much lower than what was first theorized to be
appropriate values.

Dynamicity in static detections

Static detections from reflective surfaces were initially thought to be time invari-
ant, as long as those surfaces were not moved. Previous analysis by our industry
supervisors had found that the static detections would be dominated by objects that
reflect radio waves well, like house corners or metal poles, meaning large sections
of walls would be invisible. This is due to the poor angle resolution of the radar
used, meaning it is difficult to resolve individual detections for objects in the same
range-doppler bin.
However, an interesting phenomenon was discovered when testing the mapping
algorithm. When a person moved through the scene, the reflective surface behind
the person would become visible to the radar. Only a cursory analysis was made
to find the cause of this phenomenon, but one possible explanation is that it stems
from the fact that when an object obscures a large flat surface, the gap created in the
surface makes it easier to resolve the detections next to it in the Fourier transform.

4.3 Feature density maps

In order to find the areas where real tracks are appearing, we map the spatial and
temporal density of the features described in Section 4.1. These feature density
maps are also represented using Cartesian oriented cells in a grid, which are updated
as follows:

1. Receive track in grid cell (i, j).

2. If the value of the relevant feature is true in the grid cell: Update the grid cell
with probability p = 0.5 + αs according to Equation (2.22).
If the value of the relevant feature is false in the grid cell: Update the grid cell
with probability p = 0.5 − βs according to Equation (2.22).

30
 4.3 Feature density maps

This process results in a map similar to the occupancy grid map, but with the pur-
pose of mapping common or a lack of common occurrences of certain features.

31
5
Features

Features are the properties of the tracks that were fed into the classifiers. They
were chosen in order to capture only the most relevant information in the data, this
selection process is described in Section 6.8. The features are divided into two types,
track-specific features which were extracted from the data generated by the tracker,
and scene-specific features which were extracted from the scene models.

5.1 Track-specific features

The track-specific features mentioned in this report are shown with a brief expla-
nation in Table 5.1. These features are defined for each track in each frame. The
strongest, closest, and longest features are binary, and computed by comparing the
given track with other tracks currently active in the same frame. Similarly, relative
SNR is a comparative measure, and is calculated by taking the SNR of the given
track, and dividing it with the highest SNR of all the tracks in the frame. The range
std and doppler std features are created by calculating the standard deviation of the
respective measure for all the detections in the cluster belonging to that track. Fi-
nally, the track lifetime feature is the amount of time in seconds that the tracker has
been able to keep the track alive.
Some other track-specific features that were considered are weakest, furthest
and shortest. These features can be thought of as approximate indicators for ghost
targets. As the names suggest, these features are essentially the opposites of the
other binary features strongest, closest and longest. However, when deciding which
features to use as described in Section 6.8, these features were found to not increase
performance, and was therefore not included in the final models. A possible ex-
planation for this is that a multitude of ghost tracks are often created from one real
track, and these binary features can only be active for one track at a time, essentially
missing many of the ghost tracks.

32
 5.2 Scene-specific features

Table 5.1 Explanation of track-specific features.

Feature Explanation
track lifetime The amount of time in seconds the tracker has kept the track alive
range std Standard deviation of range of detections belonging to the track
doppler std Standard deviation of Doppler of detections belonging to the track
relative SNR SNR of track divided by the highest SNR of all current tracks
strongest 1 if RCS of track is highest of all current tracks, otherwise 0
closest 1 if range of track is lowest of all current tracks, otherwise 0
longest 1 if lifetime of track is highest of all current tracks, otherwise 0

5.2 Scene-specific features

The scene-specific features were created by inputting the position of the track into
the scene model, and are listed in Table 5.2. The three density features are created
by finding the cell that represents the tracks current position in the corresponding
feature density map, and taking the value of that cell. The occupancy probability
feature is created in the same way from the occupancy grid map of the reflective
surfaces. Finally, the occlusion score feature is designed to indicate whether there
is a clear line of sight from the radar to a track, or if the track is occluded by a static
object. It is created using the occupancy grid map of the reflective surfaces and is
calculated as follows:

1. Receive track in grid cell (i, j).

2. Use Bresenhams line algorithm [Bresenham, 1965] to find all grid cells in a
straight line from cell (i, j) to the grid cell containing the radar.
3. The occlusion score is the maximum value of the occupancy probability of
these grid cells.

Table 5.2 List and explanation of scene-specific features.

Feature Explanation
occupancy probability The occupancy probability in the grid map
occlusion score Approximate probability that a track is occluded
strongest density Aggregated density of "strongest" feature
closest density Aggregated density of "closest" feature
longest density Aggregated density of "longest" feature

33
6
Model structures and
evaluation

6.1 General structure

Two types of classifier model structures were considered in this thesis. In one of the
structures, only the track-specific features from Section 5.1 were used for classifica-
tion. This model structure is visualized as in Figure 6.1. The classifier block in this
structure was implemented with a random forest in Model 1 and with a multilayer
perceptron in Model 3, both of which are described later in this section.
The other structure can be seen in Figure 6.2. In this structure, a model of
the scene is created and scene-specific features are extracted from the model as
in Section 5.2. The classifier block in this structure was implemented with a ran-
dom forest in Model 2 and with a multilayer perceptron in Model 4, both of which
are described later in this section. All machine learning models were implemented
in the Python programming language using the PyTorch framework [Paszke et al.,
2019].

6.2 Naive classifier

To be able to compare our models with a baseline model we created a simple naive
classifier. The classifier predicts a track to be a real track if the track is currently
the closest one to the radar of all active tracks in the frame, otherwise it predicts
the track to be a ghost track. This prediction is identical to the value of the feature
"closest" described in Table 5.1. This is a naive approach because if there are several
real tracks active at once, then only the closest one of the real tracks will be classified
correctly. It could also happen that the radar does not have a clear line of sight to a
target, and the closest track is in fact a ghost track caused by reflections.

34
 6.3 Model 1 — Random forest with only track-specific features

Figure 6.1 Model structure of classifier with no scene model.

6.3 Model 1 — Random forest with only track-specific

features
Model 1 was a random forest classifier, where only the track-specific features were
used. The number of decision trees used in the random forest was 100. As men-
tioned in Section 2.5, the number of features considered in the decision trees were
√
n f rounded to the nearest integer where n f is the total number of features. Be-
cause the dataset used contained more samples of ghost targets than real targets, the
samples were weighted 1 : 2 so that each sample of a real target was weighted equal
to two ghost targets when making a split in the decision trees to make the dataset
more balanced.

6.4 Model 2 — Random forest with scene-specific

features
Model 2 was a random forest classifier, where both track-specific and scene-specific
features were used. Apart from the difference in the features used, Model 2 had
exactly the same structure as Model 1.

35
Chapter 6. Model structures and evaluation

Figure 6.2 Model structure of classifier which uses a scene model to improve clas-
sification performance.

6.5 Model 3 — Multilayer perceptron with only

track-specific features
Model 3 consisted of a multilayer perceptron where only track-specific features
were used. Three hidden layers were used where each hidden layer consisted of
128 nodes. Dropout as described in Section 2.5 was used in all of the hidden lay-
ers when training the network with a dropout rate of 0.3. ReLU was used as the
activation function in all of the layers. The loss function used was the cross en-
tropy loss function described in Section 2.5. The loss function weights were 1 for
samples from ghost targets and 2 for samples from real targets, meaning that every
sample from a real target had twice as much impact on the loss as a sample from
a ghost target. The Adam optimization algorithm [Kingma and Ba, 2017] was used
when training the model. The learning rate used to update the network weights was
η = 10−4 . The network was trained for 50 epochs, meaning that on average, every
sample was used to update the network weights 50 times.

6.6 Model 4 — Multilayer perceptron with scene-specific

features
Model 4 consisted of a multilayer perceptron with exactly the same structure as
Model 3, the only difference being that scene-specific features were used in addition
to track-specific features.

36
 6.7 Evaluation

6.7 Evaluation
In order to evaluate the performance of the classifiers, leave-one-group-out cross-
validation (LOGOCV) was used. LOGOCV is a special case of leave-one-out cross-
validation [Magnusson et al., 2020], the difference being that LOGOCV omits a
group of samples in each training set instead of a single sample. The groups in
our case were the different recordings in our dataset. Data from all but one record-
ing was used to train the classifier model, and then precision, recall and F1 -score
were calculated for the two classes, as well as the weighted and unweighted aver-
ages of these metrics over the classes. This was then repeated for all the record-
ings in the dataset. Finally, the metrics were averaged over all the splits, with the
splits weighted equally regardless of the length of the recording being evaluated.
We consider this method to be the harshest but most fair method of evaluating the
classification performance.
Another method of evaluating the classification would be to split every track
in every frame randomly into a training set and evaluation set regardless of which
recording the track is from. We will refer to this as random split evaluation. This
yields a much better classification performance, but tells us little about how the
classifier would perform in a completely new setting.

6.8 Feature selection

In order to determine which features were useful for the classification task, the
models were trained on a random subset of the dataset with all potential features
included. Then, new models were trained on the same dataset where one feature at a
time was left out and the classification performance of these models were compared
to the performance of the model with all potential features included. If no significant
reduction in performance could be measured when excluding a feature, that feature
was deemed not worth including in the final models. The resulting features are the
ones described in Chapter 5. The reason for using random split evaluation here
was to make the process computationally feasible. This selection of features is not
guaranteed to give the best results when using LOGOCV evaluation, but it is an
approximation of which the most optimal features to include are.

6.9 Model parameter tuning

As a result of the computational complexity of the chosen cross-validation evalu-
ation method, a true grid search could not be performed for the model parameters
such as the structure of the neural networks and the learning rate used when training
the neural networks. A grid search is the process of training and evaluating a model
with every possible combination of some discrete set of parameter values (every
point in the parameter space). The number of decision trees used in the random

37
Chapter 6. Model structures and evaluation

forests as well as the number and size of the layers in the multilayer perceptrons
were chosen large enough such that an increase in the size of the models would
yield no significant improvements in the classification performance. To make this
feasible this was done without cross-validation, by instead using the random split
method of evaluation. Small models were trained and evaluated, iteratively increas-
ing the model sizes until no significant increase in performance could be measured.

38
7
Results

7.1 Classification results

Naive classifier
The naive classifier does not require any training data, and thus LOGOCV evalu-
ation as described in Section 6.7 is not necessary, but it was still used in order to
obtain the most fair comparison with the other classifier models. The classification
results from the naive classifier using LOGOCV evaluation can be seen in Table 7.1
The success of the naive classifier depends completely on the amount of datapoints
which only have one track in a given frame, as the naive classifier will be very ac-
curate for such datapoints. However, if multiple real tracks are frequently active at
the same time in one frame, then the naive classifier will incorrectly classify all but
one of these real tracks as ghost tracks. This would lead to a poor recall score for
real targets in the evaluation.

Table 7.1 Classification results for the naive classifier.

Class precision recall F1 score
Ghost target 0.89 0.90 0.89
Real target 0.75 0.74 0.73
weighted average 0.86 0.86 0.85

Model 1 — Random forest with only track-specific features

The classification results for the random forest with only track-specific features
using LOGOCV evaluation can be seen in Table 7.2. The class-weighted F1 score
is better than the naive classifier which is the main metric considered here when
evaluating the models. It can also be noted that the precision and recall scores are
higher for both the "Ghost target" and "Real target" classes individually.

39
Chapter 7. Results

Table 7.2 Classification results for Model 1.

Class precision recall F1 score
Ghost target 0.90 0.94 0.91
Real target 0.82 0.75 0.77
weighted average 0.88 0.89 0.88

Model 2 — Random forest with scene-specific features

The classification results using LOGOCV evaluation for the random forest with
scene-specific features included can be seen in Table 7.3. Including scene-specific
features in the random forest classifier improved the performance slightly compared
to using only track-specific features, but the improvement is quite small. A possible
explanation is that many of the recordings in the dataset were too short for the
scene-specific features to provide relevant information.

Table 7.3 Classification results for Model 2.

Class precision recall F1 score
Ghost target 0.90 0.95 0.92
Real target 0.86 0.75 0.79
weighted average 0.89 0.90 0.89

Model 3 — Multilayer perceptron with only track-specific

features
The multilayer perceptron classifier with exclusively track-specific features also
outperforms the naive classifier in F1 score using LOGOCV evaluation as can be
seen in Table 7.4. However, by noting that the recall score is high for ghost tar-
gets but low for real targets, we can conclude that this classifier is skewed towards
classifying targets as ghost targets. One possible explanation is that the dataset was
imbalanced with more ghost targets represented, and weighting real targets to be
more important was not enough to overcome this imbalance.

Table 7.4 Classification results for Model 3.

Class precision recall F1 score
Ghost target 0.86 0.98 0.91
Real target 0.91 0.63 0.73
weighted average 0.87 0.90 0.87

Model 4 — Multilayer perceptron with scene-specific features

As can be seen in Table 7.5, including scene-specific features in the multilayer per-
ceptron classifier does not improve the performance when using LOGOCV evalu-
ation. Just like in Model 3, the recall score for real targets are low, but the recall

40
 7.1 Classification results

score for ghost targets is very high. We can see that there is an interesting difference
between the classifiers that use a random forest, and the ones that use a multilayer
perceptron. The difference is that the classification results are more balanced for
Model 1 and Model 2, while Model 3 and Model 4 classify more samples as ghost
targets. A possible explanation for this is that there are a few features which have a
particularly strong correlation with the class of the sample. Due to the inbalanced
dataset, the multilayer perceptrons can transform these features such that most of
the samples get classified as ghost targets. However, a random forest consists of
many decision trees which each only uses a subset of the features. If these particu-
larly informative features are not selected when fitting a particular decision tree, the
classification result could be worse for that particular tree, essentially meaning that
the results are more random and thus more balanced.

Table 7.5 Classification results for Model 4.

Class precision recall F1 score
Ghost target 0.86 0.97 0.91
Real target 0.92 0.62 0.73
weighted average 0.87 0.89 0.87

Random split evaluation

Classification results when using random split evaluation for Model 1 and Model 2
can be seen in Tables 7.6 and 7.7, respectively. The random split evaluation method
does not say anything about how well the classifier performs on unseen data from
completely new scenes. However, by comparing the results from Model 1 and
Model 2 and noting that the performance is better for Model 2, we can conclude
that the scene-specific features contain some information that is useful for the clas-
sification task. This observation combined with the marginal performance gain from
including scene-specific features when using LOGOCV evaluation could mean that
the scene-specific features contain at least some useful information, but do not seem
to generalize well to new scenes.

Table 7.6 Classification results using the random split evaluation method for
Model 1.
Class precision recall F1 score
Ghost target 0.93 0.96 0.94
Real target 0.84 0.75 0.80
weighted average 0.91 0.91 0.91

41
Chapter 7. Results

Table 7.7 Classification results using the random split evaluation method for
Model 2.
Class precision recall F1 score
Ghost target 0.97 0.99 0.98
Real target 0.95 0.90 0.92
weighted average 0.97 0.97 0.97

7.2 Visualization of scene models

In this section we will show some examples of occupancy grid maps and feature
density maps produced in this thesis. The maps are shown with their state at the
end of the recorded scenes. All maps in this section are produced with the same
mapping parameters, which can be seen in Tables 7.9 and 7.8.
One of the scenes where data for our thesis was recorded is shown in Figure 7.1.
The radar faces the corner of a building, and there is a metal container in front and
to the left of the radar. Data was collected while we walked around in front of the
radar for a period of approximately two minutes. In the occupancy map shown in
Figure 7.2, the walls of the building, the container in the left part of the photo, and
the metal furniture in front of the radar are clearly visible. However, the map also
shows static objects beyond the walls of the building. We believe that these objects
are the results of multipath effects. Two feature density maps of the same scene can
be seen in Figure 7.3. The left image shows the density map of the closest feature,
and in this case the density map seems to show where real targets have been present
(dark area) and where ghost targets have been present (light area) in the scene. The
right image in Figure 7.3 shows the density map of the strongest feature. It appears
to show a higher density around the area where the container is located.

Figure 7.1 Photograph (left) and aerial view (right) of one of the recorded scenes.
The red triangle in the right image represents the position and orientation of the radar.
The blue rectangle represents the metal container that was missing in the original
image.

42
 7.2 Visualization of scene models

Figure 7.2 Occupancy grid map produced by recording the scene shown in
Figure 7.1. The walls of the building, the container and the metal furniture are visible
in the map.

Figure 7.3 Density maps of "closest" feature (left) and "strongest" feature (right).

Another example of a recorded scene can be seen in Figure 7.4. The scene con-
sists of a half-full parking lot with a footpath through the middle of it and some trees
next to the path. Data was collected for approximately 10 minutes, during which
some people walked in the parking lot and along the footpath in the middle of the
scene. Figure 7.5 shows the occupancy map to the left and density map of the closest
feature to the right. The reason that the occupancy map looks more sharply defined
than the occupancy map in Figure 7.2 is that the map has been created over a longer
period of time, and has thus been updated with more detections. In the occupancy
map, a problem with the occlusion feature can be seen. The cars and trees in front of
the radar resemble a wall or fence in the map, and thus all tracks appearing behind
this line of trees and cars will have a high occlusion feature value even though they
are not really occluded by a reflective surface. The left image in Figure 7.5 shows
what a feature density map might look like in a more realistic scene. The footpath
in the middle is clearly visible, where the presence of real targets is high relative to
the rest of the scene.

43
Chapter 7. Results

Figure 7.4 Photograph (left) and aerial view (right) of one of the recorded scenes.
The red triangle in the right image represents the position and orientation of the radar.

Figure 7.5 Occupancy map (left) and density map of "closest" feature (right) pro-
duced by recording the scene in Figure 7.4.

Table 7.8 Parameters for the occupancy grid map

Grid dimensions 240x120 m2
Cell dimensions 1x1 m2
p0 0.5
σr2 0.25 m
σθ2 0.75◦
αm 0.01
βm 0.0005
Ns 10

Table 7.9 Parameters for the feature density maps

Grid dimensions 240x120 m2
Cell dimensions 1x1 m2
αs 0.01
βs 0.002

44
8
Discussion

8.1 Limitations
Quality of data
A significant problem with the data collected and used in this thesis is that some
sequences do not really resemble realistic scenarios. Some sequences of data were
recorded with the specific purpose of eliciting ghost tracks, and this data can cer-
tainly be used to train classifier models that only use track-specific features. How-
ever, when building a model of the scene it is preferable to have longer and more
realistic sequences. This is because of the fact that more detections are collected
to improve the quality of the maps of the static environment and the feature den-
sity maps. Longer and more realistic sequences of data are necessary to improve
the quality of the scene-specific features created, and furthermore, to improve the
performance of the classifier models that use scene specific-features.
Another consequence of these sequences recorded with the purpose of elicit-
ing ghost tracks is that the dataset is quite imbalanced, with about three times as
many samples from ghost tracks compared to real tracks. An attempt was made to
mitigate this, with weighted loss functions and samples as described in Chapter 6.
Regardless, the classifier models that were trained in this thesis have a propensity
towards classifying samples as ghost tracks. To make a classifier more suited to
realistic scenarios, more realistic data would be needed for the training process.
One final issue with the data used in this thesis is the limitations of the pro-
prietary tracking software and the manual annotation process that are outlined in
Section 3.2. These limitations would lead to a small amount of tracks being anno-
tated incorrectly. A solution was briefly discussed where ambivalent or problematic
tracks would be pruned or split by manually altering the data. This was, however,
deemed too time-consuming for the marginal gain that would probably be achieved.

Evaluation of scene model

Evaluating and tuning the maps in the scene model from a visual standpoint only, is
far from the optimal way of maximizing their performance. A literature study pro-

45
Chapter 8. Discussion

duced some methods for evaluating grid maps, most promising being the solution
proposed by [Schwertfeger et al., 2010], however it would take quite some time to
implement. Since the purpose of this thesis is to classify ghost tracks, and not to
build as accurate a scene model as possible, it was deemed satisfactory to use its
effect on the classification as the only metric of its performance.
This, however, leads to another issue. Since the scene models are built itera-
tively over time, the quality of the scene models and thus also the scene-specific
features, will be low for all samples early in the recording. In a realistic scenario,
the radar will monitor the same environment for months, perhaps even years. As-
suming that mapping parameters are chosen so that the scene model converges after
a week, all tracks after that point will be classified using roughly the same converged
scene model. The effects of this could have been investigated by training and eval-
uating the classifiers where all scene-specific features are extracted from the final
converged scene models at the end of the recording. Additionally, some perfect and
some purposefully poor scene models could have been generated and tested in the
same way to see how much this impacted the classification results.

Scene model tuning

One issue with the proposed scene-specific features is that they require choosing the
value of many parameters. For instance, to create the map of the static environment
the value of seven parameters need to be chosen: σr , σθ , Ns , p0 , αm , βm , and the
grid map resolution. The parameters σr and σθ can be assigned reasonable values
by looking at the radar specifications, but the other five need to be determined by us-
ing them in the mapping algorithms and evaluating the results. Creating these maps
for all the recorded sequences requires significant processing time. It was there-
fore considered infeasible to attempt a grid parameter search, and the parameters
were tuned until the scene-specific features looked good visually. This method is of
course not guaranteed to provide the best model performance in the end, but it was
deemed to be the only feasible alternative considering the time and computational
resources that were available to us.

8.2 Future work

Not much previous research has been published on the problem of classifying ghost
tracks with a static radar. That meant we could not build upon previously proposed
solutions, but instead we had to propose our own solutions, prototype them and
evaluate them. Naturally there was not enough time to try all of the ideas that we
had, and so we prototyped the solutions that seemed the most promising, which in
the end resulted in our model proposals. We believe, however, that there is much
room for improvement given more time.

46
 8.2 Future work

Classifier
An area of improvement for the classifier is to better utilize the information con-
tained in the temporal evolution of the features of a track, perhaps by using a Re-
current Neural Network. Another alternative is to represent the features of a track
as spectrograms which are used as input to a convolutional neural network. Another
way of classifying tracks could be to perform a classification after the track has
been active for some fixed period of time, and then never change the classification.
As many ghost tracks are short lived, and samples with a short current lifetime are
prone to be classified as ghost tracks, this could alleviate some incorrect classifica-
tions of real tracks at the beginning of their lifetime.

Scene model
In its current implementation, the scene model uses the data to build its maps im-
mediately as it is available. However, in a real implementation, the scene model
could be built over a much longer time span than the data in this thesis has allowed.
Therefore, it would be possible to take a slower approach to mapping. One way is
to wait until a track has died, and then analysing it over its entire lifespan before
building the scene model using that information. This could potentially be used to
filter out tracks which are more uncertain from being added to the feature density
maps.
The occupancy grid map could be improved by adding a filtering step before the
scene-specific features are generated. Objects like trees and metal poles generate
a lot of static detections, but they do not necessarily block detections from people
walking behind them. An extreme example of this can be seen in Figure 7.5, where
the row of trees create what the scene model treats as a solid wall. Finding a way to
identify which static objects are obscuring objects behind them, and which are not,
would mitigate this problem. The occupancy grid maps produced could also be used
to find lines in the maps and thus more rigidly infer the position of walls and large
reflective surfaces in the scene. One could also expand on the work in [Nüßlein,
2021] to try and detect reflective surfaces by looking at the positions of ghost tracks
in the scene, and guessing where reflections have occurred. Finally, removing the
assumption that cells in the grid maps are independent could produce better maps
at a cost of more computationally complex map updates [Thrun et al., 2006].

47
9
Conclusion

The proposed classifier models prove especially effective in detecting ghost tracks,
but a significant proportion of the real tracks are misclassified as ghost tracks, es-
pecially in the case of Model 3 and Model 4. If the objective is to minimize false
alarms, this is preferable over having a low recall score for ghost tracks. Regardless,
we believe that this flaw could be mitigated by either using balanced sampling of
datapoints when training the models, or by extending the dataset with more bal-
anced data.
We believe that many of the comparative features proposed such as closest and
relative SNR provide a solid base on which many different types of classifiers can
be constructed. It is also our opinion that the scene-specific features extracted from
the scene models have the potential to improve the classification results. However,
to be able to utilize and evaluate these features, data recorded over longer periods
of time will be necessary, preferably from radar installations in real environments.
From a purely visual perspective, the occupancy grid maps produced surpris-
ingly accurate maps for many of the recorded scenes. The mapping algorithm is
quite simple and customizable, although it can be difficult to tune the parameters so
that the algorithm produces accurate maps for multiple scenes with just one set of
parameters.
We believe that this thesis shows that it is possible to reliably classify ghost
tracks with a stationary radar using our proposed method. However, there is a lot of
potential for better results with higher quality data, and the proposed improvements
to our models.

48
Bibliography

Bresenham, J. E. (1965). “Algorithm for computer control of a digital plotter”. IBM

Systems Journal 4:1, pp. 25–30. DOI: 10.1147/sj.41.0025.
Brownlee, J. (2020). A gentle introduction to cross-entropy for machine learn-
ing. https : / / machinelearningmastery . com / cross - entropy - for -
machine-learning (Accessed 2021/06/07).
Chen, V., F. Li, S.-S. Ho, and H. Wechsler (2006). “Micro-doppler effect in radar:
phenomenon, model, and simulation study”. IEEE Transactions on Aerospace
and Electronic Systems 42:1, pp. 2–21. DOI: 10.1109/TAES.2006.1603402.
Duhamel, P. and M. Vetterli (1990). “Fast Fourier transforms: a tutorial review and
a state of the art”. Signal Processing 19:4. ID: 271605, pp. 259–299.
Goodfellow, I., Y. Bengio, and A. Courville (2016). Deep Learning. http :
/ / www . deeplearningbook . org. MIT Press, Cambridge, Mass. ISBN:
9780262035613.
Gurney, K. (1997). An Introduction to Neural Networks. UCL Press, London, UK.
ISBN : 1857285034.
Hastie, T., R. Tibshirani, and J. Friedman (2008). The Elements of Statistical Learn-
ing (2nd ed.) Springer series in statistics, New York, NY. ISBN: 0-387-95284-5.
James, G., D. Witten, T. Hastie, and R. Tibshirani (2017). An Introduction to Statis-
tical Learning: with Applications in R, Corr. 7th printing 2017. Springer Texts
in Statistics, New York, NY, pp. 303–330. ISBN: 1461471370.
Kalman, R. E. (1960). “A new approach to linear filtering and prediction problems”.
Journal of Basic Engineering 82:1, pp. 35–45.
Kingma, D. P. and J. Ba (2017). Adam: a method for stochastic optimization. arXiv:
1412.6980 [cs.LG].
Kraus, F., N. Scheiner, W. Ritter, and K. Dietmayer (n.d.). “Using machine learn-
ing to detect ghost images in automotive radar”. In: 2020 IEEE 23rd Interna-
tional Conference on Intelligent Transportation Systems (ITSC) (September 20-
23, Online), pp. 1–7. DOI: 10.1109/ITSC45102.2020.9294631.

49
Bibliography

Laber, E. and L. Murtinho (2019). “Minimization of Gini impurity: NP-

completeness and approximation algorithm via connections with the k-means
problem”. English. Electronic notes in theoretical computer science 346.
Magnusson, M., A. Vehtari, J. Jonasson, and M. Andersen (2020). “Leave-one-out
cross-validation for bayesian model comparison in large data”. In: Chiappa, S.
et al. (Eds.). Proceedings of the Twenty Third International Conference on Ar-
tificial Intelligence and Statistics. Vol. 108. Proceedings of Machine Learning
Research. PMLR, pp. 341–351.
Nüßlein, A. (2021). Multipath Modelling for 60GHz Short Range Radar. Bachelor’s
Thesis LUNFMS-4050-2021. Lund Institute of Technology, Lund University,
Lund, Sweden.
Paszke, A., S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z.
Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M.
Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala
(2019). “Pytorch: an imperative style, high-performance deep learning library”.
In: Wallach, H. et al. (Eds.). Advances in Neural Information Processing Sys-
tems 32. Curran Associates, Inc., Red Hook, NY, pp. 8024–8035. URL: http:
//papers.neurips.cc/paper/9015-pytorch-an-imperative-style-
high-performance-deep-learning-library.pdf.
Prophet, R., J. Martinez, J.-C. F. Michel, R. Ebelt, I. Weber, and M. Vossiek (2019).
“Instantaneous ghost detection identification in automotive scenarios”. In: 2019
IEEE Radar Conference (RadarConf) (April 22-26, Boston, MA), pp. 1–6. DOI:
10.1109/RADAR.2019.8835603.
Rao, S. (2017). Introduction to mmwave sensing: FMCW radars. Texas Instru-
ments. URL: https://training.ti.com/sites/default/files/docs/
mmwaveSensing-FMCW-offlineviewing_0.pdf (visited on 2021-06-21).
Richards, M. A. (2005). Fundamentals of Radar Signal Processing. McGraw-Hill
Professional, New York, NY, pp. 42–50. ISBN: 0071444742. DOI: 10.1036/
0071444742.
Ryu, I.-h., I. Won, and J. Kwon (2018). “Detecting ghost targets using multi-
layer perceptron in multiple-target tracking”. Symmetry 10:1. DOI: 10.3390/
sym10010016.
Schwertfeger, S., A. Jacoff, C. Scrapper, J. Pellenz, and A. Kleiner (2010). “Eval-
uation of maps using fixed shapes: the fiducial map metric”. In: Proceedings
of the Int. Workshop on Performance Metrics for Intelligent Systems (PerMIS)
(September 28-30, Baltimore, MD), pp. 344–351. DOI: 10.1145/2377576.
2377638.
Sharma, S. (2017). Activation functions in neural networks. https : / /
towardsdatascience.com/activation-functions-neural-networks-
1cbd9f8d91d6 (Accessed 2021/06/07).

50
 Bibliography

Siciliano, B. and O. Khatib (2007). Springer Handbook of Robotics. Springer-

Verlag, Berlin, Heidelberg, pp. 1136–1150. ISBN: 354023957X.
Sim, P. (2014). Making waves: Robert Watson-Watt, the pioneer of radar. https:
/ / www . bbc . com / news / uk - scotland - tayside - central - 27393558
(Accessed 2021/02/26).
Smith, S. W. (1997). The Scientist & Engineer’s Guide to Digital Signal Processing.
California Technical Pub, San Diego, CA. ISBN: 0966017633.
Srivastava, N., G. Hinton, A. Krizhevsky, and R. Salakhutdinov (2014). “Dropout:
a simple way to prevent neural networks from overfitting”. Journal of Machine
Learning Research 15:56.
Thrun, S., W. Burgard, and D. Fox (2006). Probabilistic robotics. MIT Press, Cam-
bridge, Mass. ISBN: 0262201623.

51

Document name
/XQG8QLYHUVLW\
0$67(5¶67+(6,6
'HSDUWPHQWRI$XWRPDWLF&RQWURO Date of issue
%R[ -XQH
6(/XQG6ZHGHQ Document Number
7)57
Author(s) Supervisor
$QWRQ6HGLQ 6HEDVWLDQ+HXQLVFK$[LV&RPPXQLFDWLRQV$%
6ZHGHQ
'DYLG:DGPDUN $UDV3DSDGHOLV$[LV&RPPXQLFDWLRQV$%6ZHGHQ
%M|UQ2ORIVVRQ'HSWRI$XWRPDWLF&RQWURO/XQG
8QLYHUVLW\6ZHGHQ
5ROI-RKDQVVRQ'HSWRI$XWRPDWLF&RQWURO/XQG
8QLYHUVLW\6ZHGHQ H[DPLQHU 

Title and subtitle

*KRVWWDUJHWFODVVLILFDWLRQXVLQJVFHQHPRGHOVLQUDGDU
Abstract

,QVXUYHLOODQFHFRQWH[WVUDGDUVFDQEHXVHGWRPRQLWRUDQDUHDGHWHFWLQJDQGWUDFNLQJPRYLQJREMHFWV
LQVLGHLW0RQLWRUHGDUHDVLQXUEDQHQYLURQPHQWVRIWHQFRQWDLQPDQ\VXUIDFHVWKDWUHIOHFWUDGDUZDYHV
ZKLFKFDQKDYHWKHXQGHVLUHGFRQVHTXHQFHRIDVLQJOHREMHFWSURGXFLQJPXOWLSOHWUDFNVGXHWR
PXOWLSDWKSURSDJDWLRQHIIHFWV7KLVWKHVLVFRQVLGHUVDPHWKRGRILGHQWLI\LQJLIDWUDFNLVSURGXFHGE\
DUHDOREMHFWRULILWVWHPVIURPPXOWLSDWKHIIHFWV7KHSURSRVHGPHWKRGZRUNVE\FUHDWLQJD
PDFKLQHOHDUQLQJEDVHGFODVVLILHUDQGPRGHOOLQJWKHPRQLWRUHGVFHQHRYHUWLPH7UDFNVDUHDVVLJQHG
IHDWXUHVEDVHGRQWKHLUFKDUDFWHULVWLFVDQGWKHVWDWHRIWKHVFHQHPRGHOLQUHJDUGVWRWKHLUSRVLWLRQ
7KHVHIHDWXUHVDUHWKHQXVHGDVLQSXWVWRWKHFODVVLILHUPRGHOWRSURGXFHWKHFODVVLILFDWLRQ:H
SURSRVHIRXUPDFKLQHOHDUQLQJEDVHGFODVVLILHUPRGHOVZLWKWZRGLIIHUHQWVHWVRIVWUXFWXUHVDQG
IHDWXUHVXVHG7KHFODVVLILHUPRGHOVDUHFRPSDUHGWRDQDLYHFODVVLILHUPRGHOIRUUHIHUHQFH

7KHSURSRVHGPRGHOVDOORXWSHUIRUPWKHQDLYHFODVVLILHUDOWKRXJKVRPHRIWKHPDUHELDVHG$VIRU
WKHXVHIXOQHVVRIWKHVFHQHPRGHOWKHUHVXOWVDUHPL[HGEXWVKRZSURPLVH:HEHOLHYHWKDWWKHVFHQH
PRGHOFDQLPSURYHFODVVLILFDWLRQSHUIRUPDQFHIXUWKHUZLWKPRUHDQGEHWWHUGDWD

Keywords
UDGDUVXUYHLOODQFHPXOWLSDWKPDFKLQHOHDUQLQJFODVVLILFDWLRQ
Classification system and/or index terms (if any)

Supplementary bibliographical information

ISSN and key title ISBN
 
Language Number of pages Recipient’s notes
(QJOLVK 
Security classification

KWWSZZZFRQWUROOWKVHSXEOLFDWLRQV