Professional Documents
Culture Documents
Sensor Fusion For Automotive Applications
Sensor Fusion For Automotive Applications
No. 1409
Christian Lundquist
Linköping 2011
Cover illustration: The intensity map describes the density of stationary targets
along the road edges. A photo of the the driver’s view in the green car is shown
in Figure 5a on Page 216.
Christian Lundquist
[email protected]
www.control.isy.liu.se
Division of Automatic Control
Department of Electrical Engineering
Linköping University
SE–581 83 Linköping
Sweden
v
Populärvetenskaplig sammanfattning
Dagens bilar blir säkrare för varje år. Tidigare var det den passiva säkerheten med
bilbälten, krockkuddar och krockzoner som förbättrades. Nu sker de snabbaste
förändringarna inom aktiv säkerhet, med förarstödsystem såsom antisladd och
nödbromssystem och förarvarningssystem för farliga filbyten och framförvaran-
de hinder.
Förarstödsystem kräver omvärldsuppfattning, och de sensorer för detta som finns
i bruk idag är kamera och radar. Kameran ser t.ex. filmarkeringar och kan använ-
das för att detektera fordon och fotgängare. Radarn är mycket bra på att detekte-
ra rörliga objekt och på avståndsbedömning, och används idag för att följa bilen
framför. Radarn ger en mängd information som idag är outnyttjad, och ett syfte
med avhandlingen är att undersöka hur radarinformationen kan användas fullt
ut i framtida säkerhetssystem.
Ett bidrag i avhandlingen är att använda radarns mätningar av stillastående ob-
jekt på och jämte vägen för att bygga upp en lokal karta. Kartan visar områden
som är körbara och hinder som ska undvikas. Ett annat bidrag är att sortera alla
de träffar från rörliga objekt som radarn registrerar och ge en noggrann karta över
hur många andra fordon det finns, samt deras positioner, hastigheter och storle-
kar. Tillsammans bildar dessa två kartor en lägesbild som kan användas i nästa
generations kollisionsundvikande system som kan kombinera broms och styring-
repp för att på ett intelligent, säkert och kontrollerat sätt väja undan i kritiska
nödsituationer.
En annan typ av säkerhetssystem är varningssystem till föraren på förändring-
ar i fordonsdynamiska parametrar som kan försämra köregenskaperna. Exempel
på sådana parametrar är däcktryck, friktion och fellastning. Ett bidrag i avhand-
lingen är ett nytt sätt att skatta en sänkning i däcktryck, genom att kombinera
sensorer i fordonet med satellitnavigering.
De metoder som presenteras i avhandlingen har utvärderats på verkliga data från
bland annat motorvägar och landsvägar i Sverige.
vii
Acknowledgments
It all started in the end of the 80s, when my teacher asked me what I wanted to
become when I grow up. I had only one wish, to study at Chalmers; followed
by my answer my teacher laughed and told me “you will never make it, you are
too stupid". This incentivized me, and encourage by other teachers I managed to
finish school and be admitted at Chalmers. Several years later I’m now here and
I have finished my thesis, something I would never have managed alone, without
the support of many wonderful persons.
I would like to thank my supervisor professor Fredrik Gustafsson for his positive,
inspiring, encouraging and relaxed attitude and my co-supervisor Dr. Thomas B.
Schön for giving me a speedy start into academic research. Whenever I run into
problems Dr. Umut Orguner, who possesses an enormous amount of knowledge
about everything beyond least squares, pushed me in the right direction. Thank
you for your help and for always having time. Further, I would like to acknowl-
edge professor Svante Gunnarsson, who is a skillful head of the group, and his
predecessor professor Lennart Ljung. They have created a wonderful research
atmosphere which makes enjoyable going to office.
Part of the work has been performed with colleagues. I would like to mention Lic.
Karl Granstöm, since one can’t find a better partner to collaborate with. Further
I like to mention Dr. Lars Hammarstrand, at Chalmers, with whom I had very
useful and interesting discussions. Finally, I would like to thank Dr. Emre Özkan,
who opened the Gaussian window and let me see other distributions.
Andreas Andersson at Nira Dynamics AB and Dr. Andreas Eidehall at Volvo Per-
sonvagnar AB have rescued me from the simulation swamp, by supporting me
with measurement data from prototype vehicles. Dr. Gustaf Hendeby and Dr.
Henrik Tidefelt helped me with latex issues, without their support this thesis
would not be in this professional shape. Ulla Salaneck, Åsa Karmelind and Ninna
Stensgård have helped with many practical things during the years in the group.
An important part of the PhD studies is the time spent outside the office bunker,
i.e., in the fika room and in pubs. I would like to thank Lic. Jonas Callmer, Lic.
Karl Granström and Lic. Martin Skoglund for sharing vagnen with me. Further
I would like to thank Lic. Christian Lyzell, Dr. Ragnar Wallin, Dr. Henrik Ohls-
son Lic. Zoran Sjanic and Sina Khoshfetrat Pakazad for being generous and good
friends.
Dr. Wolfgang Reinelt became a mentor for me, he supported and encouraged me
during the time at ZF Lenksysteme GmbH. With him I wrote my first publica-
tions. My former boss Gerd Reimann taught me about vehicle dynamics and the
importance of good experiments; I will always remember Sinuslenken.
One of the most inspiring persons I met is Dr. Peter Bunus. Together with pro-
fessor Fredrik Gustafsson, Dr. David Törnqvist, Lic. Per Skoglar and Lic. Jonas
Callmer we started SenionLab AB. I’m looking forward to keeping on working
with all of them in the near future.
ix
x Acknowledgments
I would like to acknowledge the supported from the SEnsor fusion for Safety
(sefs) project within the Intelligent Vehicle Safety Systems (ivss) program and
the support from the Swedish Research Council under the frame project grant
Extended Target Tracking.
I would never have been able to fulfill my wish to study without the support from
my family, for this I am endlessly thankful. Finally, I would like to thank amore
della mia vita Nadia, who has brought so much love and joy into my life.
Notation xvii
I Background
1 Introduction 3
1.1 Sensor Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Automotive Sensor Fusion . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Sensor Fusion for Safety . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Extended Target Tracking . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Components of the Sensor Fusion Framework . . . . . . . . . . . . 7
1.6 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.7 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.8 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.8.1 Outline of Part I . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.8.2 Outline of Part II . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Estimation Theory 31
3.1 Static Estimation Theory . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.1 Least Squares Estimator . . . . . . . . . . . . . . . . . . . . 33
3.1.2 Probabilistic Point Estimates . . . . . . . . . . . . . . . . . . 35
3.2 Filter Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2.1 The Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.2 The Extended Kalman Filter . . . . . . . . . . . . . . . . . . 38
3.2.3 The Unscented Kalman Filter . . . . . . . . . . . . . . . . . 39
3.2.4 The Particle Filter . . . . . . . . . . . . . . . . . . . . . . . . 41
xi
xii CONTENTS
4 Target Tracking 45
4.1 Single Target Tracking . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.2 Extension to Multitarget Tracking . . . . . . . . . . . . . . . . . . . 47
4.2.1 Data Association . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2.2 Track Management . . . . . . . . . . . . . . . . . . . . . . . 50
4.3 Extended Target Tracking . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.1 Point Features . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.2 Spatial Distribution . . . . . . . . . . . . . . . . . . . . . . . 53
4.3.3 Elliptical Shaped Target . . . . . . . . . . . . . . . . . . . . 54
4.3.4 Curved Target . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.3.5 Extended Target Tracking and PHD filter . . . . . . . . . . 56
6 Concluding Remarks 77
6.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.2 Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Bibliography 81
II Publications
A Situational Awareness and Road Prediction for Trajectory Control Ap-
plications 97
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
2 Modeling the Environment with a Map . . . . . . . . . . . . . . . . 100
3 Feature Based Map . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.1 Radar and Laser . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.2 Cameras and Computer Vision . . . . . . . . . . . . . . . . 104
4 Road Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.1 Road Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.2 Mapping of the Road Lanes . . . . . . . . . . . . . . . . . . 110
4.3 Mapping of the Road Edges . . . . . . . . . . . . . . . . . . 114
CONTENTS xiii
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
Index 307
Notation
Operators
Notation Meaning
min minimize
max maximize
arg min A the value x that minimizes A
x
arg max A the value x that maximizes A
x
E expectation
Cov covariance
Var variance
diag(a, b) diagonal matrix with elements a and b
blkdiag(A, B) Block diagonal matrix with matrices A and B
Pr(A) probability of event A
AT transpose of matrix A
Std standard deviation
Tr(A) trace of matrix A
|x| absolute value of x (or cardinality if x is a set)
x̂ estimate of the stochastic variable x
, equal by definition
∼ is distributed according to
∝ proportional to
∈ belongs to
ẋ time derivative of x
xvii
xviii Notation
Notation Meaning
iW inverse Wishart distribution
D intesity or phd function
N normal or Gaussian distribution
NiW normal inverse Wishart distribution
Pois Poisson ditribution
St student-t distribution
U uniform distribution
p(x) density function of x
p(x, y) joint density function of x and y
p(x|y) conditional density function of x given y
Sets
Notation Meaning
N set of natural numbers
M set of map variables or features
R set of real numbers
S set of measurement generating points
X set of state variables
Z set of measurements
Nx number of elements in the set X
∅ empty set
Random Variables
Notation Meaning
x state
y measurement (no data association needed)
z measurement (data association needed)
u input
P state covariance
w process noise
Q process noise covariance
e measurement noise
R measurement noise covariance
m map state
p point feature
s state of a measurement generating point
θ parameter
nx dimension of the vector x
x̂k|κ estimate of x at time given measurements y1:κ
Notation xix
Notation Meaning
c Cartesian representation
p polar representation
x Cartesian position x-coordinate (longitudinal)
y Cartesian position y-coordinate (lateral)
z Cartesian position z-coordinate (vertical)
w width
l length
a acceleration
v velocity
u, v pixel coordinates
d displacement vector
r range (between sensor and target)
ψ heading angle, yaw angle or bearing (angle around z-
axis)
Miscellaneous
Notation Meaning
k discrete time variable
T sample time
lb wheel base (distance between front and rear axle)
lt wheel track (distance between left and right wheel)
co road curvature
c1 road curvature derivative
xx Notation
Abbreviations
Abbreviation Meaning
abs antilock braking system
acc adaptive cruise control
c2c car to car communication
can controller area network
eiv errors in variables
eio errors in output
ekf extended Kalman filter
ela emergency lane assist
et-gm-phd extended target Gaussian mixture probability hypoth-
esis density
fisst finite set statistics
fmcw frequency modulated continuous-wave (a radar sys-
tem)
gm-phd gaussian mixture probability hypothesis density
gnn global nearest neighbor (data association)
gps global positioning system
imu inertial measurement unit
kf Kalman filter
lka lane keeping assistance
ls least Squares
mae mean absolute error
map maximum a posteriori
mc Monte Carlo (experiments based on repeated random
sampling)
mgp measurement generating point (reflection point on tar-
get)
ml maximum likelihood
mpf marginalized particle filter
nn nearest neighbor (data association)
ogm occupancy grid map
olr optical lane recognition
ospa optimal subpattern assignment (multitarget evalua-
tion measure)
pdf probability density function
phd probability hypothesis density
pf particle Filter
rfs random finite set
rmse root mean square error
slam simultaneous localization and mapping
s.t. subject to
ukf unscented Kalman filter
virt virtual sensor measurement
wls weighted least squares
Part I
Background
Introduction
1
This thesis is concerned with the problem of estimating the motion of a vehicle
and the characteristics of its surroundings. More specifically, the description of
the ego vehicle’s surroundings consists of other vehicles and stationary objects
as well as the geometry of the road. The signals from several different sensors,
including camera, radar and inertial sensor, must be combined and analyzed to
compute estimates of various quantities and to detect and classify many objects
simultaneously. Sensor fusion allows the system to obtain information that is
better than if it was obtained by individual sensors.
Situation awareness is the perception of environmental features, the comprehen-
sion of their meaning and the prediction of their status in the near future. It
involves being aware of what is happening in and around the vehicle to under-
stand how the subsystems impact on each other.
Sensor fusion is introduced in Section 1.1 and its application within the automo-
tive community is briefly discussed in Section 1.2. The study presented in this
thesis was conducted within two Swedish research project, briefly described in
Section 1.3 and 1.4. The sensor fusion framework and its components, such as
infrastructure, estimation algorithms and various mathematical models, are all
introduced in Section 1.5. Finally, the chapter is concluded with an overview
of the author’s publications in Section 1.6, a statement of the contributions in
Section 1.7 and the outline of this thesis given in Section 1.8.
3
4 1 Introduction
Measurement Model
Figure 1.1: The main components of the sensor fusion framework are shown
in the middle box. The framework receives measurements from several sen-
sors, fuses them and produces one state estimate, which can be used by sev-
eral applications.
in some sense better than it would be if the sensors were used individually. The
term better can in this case mean more accurate, more reliable, more available
and of higher safety integrity. Furthermore, the resulting estimate may in some
cases only be possible to obtain by using data from different types of sensors.
Figure 1.1 illustrates the basic concept of the sensor fusion framework. Many
systems have traditionally been stand alone systems with one or several sensors
transmitting information to only one single application. Using a sensor fusion
approach it might be possible to remove one sensor and still perform the same
tasks, or add new applications without the need to add new sensors.
Sensor fusion is required to reduce cost, system complexity and the number of
components involved and to increase accuracy and confidence of sensing.
the controller area network (can). It is a serial bus communication protocol de-
veloped by Bosch in the early 1980s and presented by Kiencke et al. (1986) at the
sae international congress in Detroit. An overview of the can bus, which has be-
come the de facto standard for automotive communication, is given in Johansson
et al. (2005).
As already mentioned, the topic of this thesis is how to estimate the state vari-
ables describing the ego vehicle’s motion and the characteristics of its surround-
ings. The ego vehicle is one subsystem, labeled (E) in this work. The use of
data from the vehicle’s actuators, e.g., the transmission and the steering wheel,
to estimate a change in position over time is referred to as odometry. The ego
vehicle’s surroundings consists of other vehicles, referred to as targets (T), and
stationary objects as well as the shape and the geometry of the road (R). Map-
ping is the problem of integrating the information obtained by the sensors into a
given representation, denoted (M), see Adams et al. (2007) for a recent overview
and Thrun (2002) for an older survey. Simultaneous localization and mapping
(slam) is an approach used by autonomous vehicles to build a map while at the
Camera Radar
Detects other vehicles, lane other vehicles, sta-
markings, pedestri- tionary objects
ans
Classifies objects yes no
Azimuth angle high accuracy medium accuracy
Range low accuracy very high accuracy
Range rate not very high accuracy
Field of View wide narrow
Weather Condi- sensitive to bad visi- less sensitive
tions bility
6 1 Introduction
(a) (b)
Figure 1.2: Figure (a) shows the camera and Figure (b) the front looking
radar in a Volvo S60 production car. Courtesy of Volvo Car Corporation.
same time keeping track of their current locations, see e.g., Durrant-Whyte and
Bailey (2006); Bailey and Durrant-Whyte (2006). This approach is not treated in
this thesis.
ses (Bengtsson, 2008; Danielsson, 2008; Lundquist, 2009) have been produced.
An overview of the main results in the project is given in Ahrholdt et al. (2009)
and the sensor fusion framework is well described in Bengtsson and Danielsson
(2008). Furthermore it is worth mentioning some of the publications produced by
the project partners. Motion models for tracked vehicles are covered in Svensson
and Gunnarsson (2006); Gunnarsson et al. (2006); Sörstedt et al. (2011). A bet-
ter sensor model of the tracked vehicle is presented in Gunnarsson et al. (2007).
Detection of lane departures and lane changes of leading vehicles is studied in
Schön et al. (2006), with the goal to increase the accuracy of the road geometry es-
timate. Computational complexity for systems obtaining data from sensors with
different sampling rates and different noise distributions is studied in Schön et al.
(2007).
Paper D in this thesis describes a method to estimate and represent stationary
objects along the road edges. This publication is based on data collected from the
sefs prototype car and the work was carried out within the project.
model based methods. The systems discussed in this thesis are primarily dy-
namic and they are modeled using stochastic difference equations. More specif-
ically, the systems are modeled using the discrete-time nonlinear state space
model
xk+1 = fk (xk , uk , wk , θk ), (1.1a)
yk = hk (xk , uk , ek , θk ), (1.1b)
where (1.1a) describes the evolution of the state variable x over time and (1.1b)
explains how the state variable x relates to the measurement y1 . The state vector
at time k is denoted by xk ∈ Rnx , with elements x1 , . . . , xnx being real numbers.
Sensor observations collected at time k are denoted by yk ∈ Rny , with elements
y1 , . . . , yny being real numbers. The model fk in (1.1a) is referred to as the pro-
cess model , the motion model , the dynamic model or the system model, and it
describes how the state propagates in time. The model hk in (1.1b) is referred to
as the measurement model or the sensor model and it describes how the state
is propagated into the measurement space. The random vector wk describes the
process noise, which models the fact that the actual state dynamics is usually
unknown. The random vector ek describes the sensor noise. Furthermore, uk
denotes the deterministic input signals and θk denotes the possibly unknown
parameter vector of the model.
The ego vehicle constitutes an important dynamic system in this thesis. The yaw
and lateral dynamics are modeled using the so called single track model. This
model will be used as an example throughout the thesis. Some of the variables
and parameters in the model are introduced in Example 1.1.
1.1 Example: Single Track Ego Vehicle Model
A so called bicycle model is obtained if the wheels at the front and the rear axle of
a passenger car are modeled as single wheels. This type of model is also referred
to as a single track model and a schematic drawing is given in Figure 1.3. Some
examples of typical variables and parameters are:
State variables x: the yaw rate ψ̇E and the body side slip angle β, i.e.,
h iT
x = ψ̇E β . (1.2)
Measurements y: the yaw rate ψ̇E and the lateral acceleration ay , i.e.,
h iT
y = ψ̇E ay , (1.3)
which both are measured by an inertial measurement unit (imu).
Input signals u: the steering wheel angle δs , which is measured with an angu-
lar sensor at the steering column and the velocity v̇, which is measured by
1 Note that the measurement vector is denoted y when the sensor measures one or several specific
quantities, e.g., acceleration and yaw rate from an imu sensor. However, when the number of sensor
measurements (observations) depends on the environmental circumstances the measurement vector
is denoted z, e.g., for a radar which observes an unknown number of targets. This usually leads to a
data association problem.
1.5 Components of the Sensor Fusion Framework 9
ρ αf
δf
v
β
ψE y W
CoG
αr
OW x
Figure 1.3: Illustration of the geometry for the single track model, describing
the motion of the ego vehicle. The ego vehicle velocity vector v is defined
from the center of gravity (CoG) and its angle to the longitudinal axis of the
vehicle is denoted by β, referred to as the body side slip angle. Furthermore,
the slip angles are referred to as αf and αr . The front wheel angle is denoted
by δf and the current driven radius is denoted by ρ.
Parameters θ: the vehicle mass m, which is weighed before the tests, the steering
ratio is between the steering wheel angle and the front wheels, which has to
be estimated in advance, and the tire parameter Cα , which is estimated on-
line, since the parameter value changes due to different road and weather
conditions.
The model (1.1) must describe the essential properties of the system, but it must
also be simple enough to be efficiently used within a state estimation algorithm.
Chapter 3 describes algorithms that are used to compute estimates of the state xk
and the parameter θk in (1.1).
Before describing the individual steps of the sensor fusion framework another
important example is presented in Example 1.2.
The different steps of a typical sensor fusion algorithm, as the central part of
the larger framework, are shown in Figure 1.4. The algorithm is initiated using
a prior guess of the state x0 or, if it is not the first iteration, the state estimate
x̂k−1|k−1 from the previous time step k − 1 is used. New measurements Zk are col-
lected from the sensors and preprocessed at time k. Model (1.1) is used to predict
the state estimate x̂k|k−1 and the measurement ẑk|k−1 . For Example 1.2 it is nec-
essary to associate the radar observations Zk with the predicted measurements
Zk|k−1 of the existing state estimates and to manage the tracks, i.e., initiate new
b
states and remove old, invalid states. The data association and track management
are further discussed in Section 4.2. Finally, the new measurement zk is used to
calculate the state estimate x̂k|k at time k in the so called measurement update
step. The prediction and measurement update are described in Section 3.2. This
algorithm is iterated, x̂k|k is used to predict x̂k+1|k , new measurements Zk+1 are
collected at time k + 1 and so on. The state estimation theory, as part of the sensor
fusion framework, is discussed further in Chapter 3. Note that in Example 1.1,
the data association and track management are obviously not needed, since there
the data association is assumed fixed. In the example the measurements y from
the pre-processing are fed directly to the measurement update.
1.6 Publications
Published work of the author that are of relevance to this thesis are listed below.
The publications are clustered in groups preambled by a short summary. The
publications are principally listed in chronological order.
The author has been involved in the development of an active steering system
prior to starting his PhD-studies. The active steering system superimposes an
1.6 Publications 11
Zk pre-
sensor
processing
Zk
Figure 1.4: The new measurements Zk contain new information and are as-
sociated to the predicted states b
Xk|k−1 and thereafter used to update them to
obtain the improved state estimates b Xk|k .
electronically controlled angle to the driver’s steering input. The aim of the sys-
tem is to reduce the driver’s steering efforts at low velocities, to reduce the vehi-
cle’s sensibility at very high velocities and to intervene when the vehicle tends to
become unstable. The system is thoroughly described in
W. Reinelt and C. Lundquist. Mechatronische Lenksysteme: Mod-
ellbildung und Funktionalität des Active Front Steering. In R. Is-
ermann, editor, Fahrdynamik Regelung - Modellbildung, Fahrassis-
tenzsysteme, Mechatronik, pages 213–236. Vieweg Verlag, September
2006a.
W. Reinelt, W. Klier, G. Reimann, C. Lundquist, W. Schuster, and R.
Großheim. Active front steering for passenger cars: System modelling
and functions. In Proceedings of the IFAC Symposium on Advances
in Automotive Control, Salerno, Italy, April 2004.
Sensor fusion based monitoring systems are described in
W. Reinelt, C. Lundquist, and H. Johansson. On-line sensor monitor-
ing in an active front steering system using extended Kalman filtering.
In Proceedings of the SAE World Congress, SAE paper 2005-01-1271,
Detroit, MI, USA, April 2005.
W. Reinelt and C. Lundquist. Observer based sensor monitoring in
an active front steering system using explicit sensor failure modeling.
In Proceedings of the IFAC World Congress, Prague, Czech Republic,
July 2005.
S. Malinen, C. Lundquist, and W. Reinelt. Fault detection of a steering
wheel sensor signal in an active front steering system. In Preprints
of the IFAC Symposium on SAFEPROCESS, pages 547–552, Beijing,
China, August 2006.
C. Lundquist and W. Reinelt. Electric motor rotor position monitor-
ing method for electrically aided steering system e.g. steer by wire,
12 1 Introduction
A marginalized particle filter based method to estimate the tire radii, and thereby
being able to detect pressure losses is presented in
1.7 Contributions
The main contributions of this thesis are briefly summarized and presented be-
low:
Mapping: An overview of different methods to map stationary objects in the sur-
roundings of the vehicle are summarized in Paper A.
Road curvature estimation: The estimation from camera and radar is improved
by using information about the ego vehicle motion. The results are pre-
sented in Paper B.
Road edge estimation: Errors in variables methods are applied in order to track
the road edges using point measurements, and the results are given in Pa-
per C. The edges are modeled as extended targets.
Intensity based map: An intensity function is used to represent stationary ob-
jects along the road. Structures in the road are used to improve the update
and representation of the map as described in Paper D.
Extended target tracking: A modification of the gm-phd filter, which is able to
handle extended targets, is presented in Paper E.
Target size estimation: An approach to estimate the size and the shape is intro-
duced in Paper F. The tracking framework contains a hybrid state space
where measurement generating points and the measurements are modeled
by random finite sets and target states by random vectors.
Tire radii estimation: The marginalized particle filter is applied to the problem
of estimating the tire radii of a vehicle. The Bayesian method, presented in
Paper G, is efficient and the estimate is more accurate than estimates from
comparable methods.
21
22 2 Models of Dynamic Systems
Road models are treated in Paper A to D. A good overview of available and com-
monly used road models is given in Paper A, and therefore no introduction to
road modeling in general is given in this chapter. However, road models are of-
ten combined with ego vehicle models and described in one single state space
2.1 Overview of the Models Used in the Thesis 23
Table 2.1: Typical ranges for the vehicle parameters used in the single track
model.
m Izz Cα lf + lr
kg kgm2 N/rad m
model. One approach is presented in Paper B and compared with two other vehi-
cle models with road interaction, which are described as follows.
The first model is described in Example 2.2 and it is commonly used for au-
tonomous driving and lane keeping. This model is well described by e.g., Dick-
manns (2007) and Behringer (1997). Note that the ego vehicle’s motion is mod-
eled with respect to a road fixed coordinate frame, unlike the single track model
in Example 2.1, which is modeled in a Cartesian world coordinate frame.
The steering wheel angle might have a bias, for example if the sensor is not cal-
ibrated, which leads to an accumulation of the side slip angle β in (2.9). Other
reasons for a steering wheel angle bias is track torsion or strong side wind, which
the driver compensates for with the steering wheel. The problem is solved by
24 2 Models of Dynamic Systems
Note that the curvature c0 is included in (2.6). The curvature c0 is usually treated
as a state variable and in this example the vehicle motion model is augmented
with the simple clothoid road model, see Paper A.
Another and simpler vehicle model is obtained if the side slip angle is omitted
and the ego vehicle’s yaw rate ψ̇E is used instead of the steering wheel angle. The
model is summarized in Example 2.3, and thoroughly described together with
results in Eidehall (2007); Eidehall et al. (2007); Eidehall and Gustafsson (2006);
Gern et al. (2000, 2001); Zomotor and Franke (1997).
More advanced vehicle models with more degrees of freedom, including the two
track model, are described by Schofield (2008).
2.1 Overview of the Models Used in the Thesis 25
In this work, only measurements from the ego vehicle’s sensors are available; that
is the target’s motion is measured using the ego vehicle’s radar and camera. This
is the reason for why the target model is simpler than the ego vehicle model. The
targets play an important role in the sensor fusion framework presented in this
work, but little effort has been spent modeling their motion. Instead standard
models from target tracking literature are used. A survey of different process
models and measurement models are given by Rong Li and Jilkov (2003) and
Rong Li and Jilkov (2001), respectively. The subject is also covered in the books
by Blackman and Popoli (1999) and Bar-Shalom et al. (2001). One typical target
model is given in Example 2.4. This model is used in the Papers F and G, and a
somewhat similar model, called constant velocity model, is used in Paper E.
Simpler than using the exact sampling formula (2.19) is it to make use of the
standard forward Euler method, which approximates (2.2a) according to
xk+1 ≈ xk + T a(xk , uk , wk , θk ) , fk (xk , uk , wk , θk ). (2.21)
This is a very rough approximation with many disadvantages, but it is frequently
used due to its simplicity. Example 2.6 shows the Euler approximation of the
coordinated turn model.
The model is linear in the input δf ,k . However, the input vk is implicitly modeled
in the matrices Fk (vk , θk ), Gku (vk , θk ) and Hk (vk , θk ).
Linear state space models and linear system theory in general are thoroughly de-
scribed by Rugh (1996) and Kailath (1980).
In Example 1.1 an ego vehicle model was introduced, where the steering wheel
angle and the vehicle velocity were modeled as deterministic input signals. This
consideration can be motivated by claiming that the driver controls the vehicle’s
lateral movement with the steering wheel and the longitudinal movement with
the throttle and brake pedals. Furthermore, the steering wheel angle and the
velocity are measured with less noise than the other measurement signals, and
2.4 Nonlinear State Space Model with Additive Noise 29
they are often pre-processed to improve the accuracy and remove bias. With
these arguments the resulting model, given in Example 2.1, may be employed.
The model is in some sense simpler than the case when these two signals are
assumed to be stochastic measurements, as shown in Example 2.8.
where T is the sample time. The first two rows of the process and measurement
models are the discretized versions of the model given in (2.4). In the second row
in the measurement model the simplification cos δf ≈ 1 and sin δf ≈ 0 is made
to not let the measurement model be dependent on the measurement δf . The
third measurement signal is the steering wheel angle δs , but the third state is the
front wheel angle δf . A possible measurement model hδf will be discussed in
Example 3.1. Finally, a random walk is assumed for the front wheel angle δf and
the velocity v in the process model.
Another way to represent the state space model is given by considering the prob-
ability density function (pdf) of different signals or state variables of a system.
The transition density p(xk+1 |xk ) models the dynamics of the system and if the
process noise is assumed additive, the transition model is given by
p(xk+1 |xk ) = pw (xk+1 − f (xk , uk , θk )), (2.27)
where pw denotes the pdf of the process noise w. The density describes the state
at time k + 1 given that information about the state at the previous time step k is
known. A fundamental property of the process model is the Markov property,
p(xk+1 |x1 , . . . , xk ) = p(xk+1 |xk ). (2.28)
This means that the state of the system at time k contains all necessary informa-
tion about the past, which is needed to predict the future behavior of the system.
30 2 Models of Dynamic Systems
31
32 3 Estimation Theory
For various reasons some systems are only modeled by a likelihood function. Of-
ten these systems are static and there exists no Markov transition density. How-
ever, most systems in this thesis are modeled using both a prediction and a like-
lihood function. In system identification, the model parameter is estimated with-
out physically describing the parameter’s time dependency, hence static estima-
tion theory is used. The state can be estimated in more or less the same way.
However, the process model (1.1a) is often given and its time transition informa-
tion is exploited to further improve the state estimate.
The origins of the estimation research field can be traced back to the work by
Gauss in 1795 on least squares (Abdulle and Wanner, 2002) and Bayes (1763)
on conditional probabilities. Bayes introduced an important theorem which has
come to be referred to as Bayes’ theorem,
p(y|x, θ)p(x, θ)
p(x, θ|y) = , (3.2)
p(y)
with which it is possible to calculate the posterior probability p(x, θ|y) given a
prior probability p(x, θ) and the likelihood function p(y|x, θ). Note that both
the measurement and the state or parameter are treated as random variables.
Another view of the estimation problem was introduced by Fisher (1922), who
claimed that the probability of an estimate should be seen as a relative frequency
of the state or parameter, given data from long-run experiments. Fisher also
treats the measurement as a random variable. The main difference to Bayes’ ap-
proach is that in Fisher’s approach there is a true state or parameter which is
treated as deterministic, but unknown. To accentuate the different views, the
likelihood is often written using `(x, θ) to emphasize that the likelihood is re-
garded as a function of the state x and the parameter θ.
After this brief historical background, the remainder of this chapter is outlined
as follows. In Section 3.1, static estimation methods based on both Fishers and
Bayes theories, are discussed. These methods can be used for both state and
parameter estimation. In Section 3.2, dynamic estimation methods are discussed.
These methods are, within the scope of this thesis, only used for state estimation
and are based solely on Bayes’ theories.
How to separate a typical estimation problem into these two parts is shown Ex-
ample 3.2.
General estimation techniques are covered by most textbooks on this topic, e.g.,
Kay (1993); Kailath et al. (2000); Ljung (1999). There are many estimation meth-
ods available, however, in this section the focus is on the methods used in Part II
of this thesis.
3.2 Example: Parameter and State Estimation
Consider the linear single track model in Example 2.7. Suppose that the state
variables are measured with external and highly accurate sensors. The yaw rate
is measured with an extra imu and the body side slip angle β is measured with
a so called Correvit® sensor, which uses correlation technology to compute the
optical flow. The vehicle in Figure 2 on Page 297 is equipped with this type of
sensor at the front. Now, the parameter θ can be estimated, according to (3.3a).
This approach has been used to find some of the vehicle parameters in Papers B
and G.
Conversely, if θ is known and y is measured, the state variables x can be estimated
using (3.3b).
This section covers estimation problems without any process model f ( · ), where
a set of measurements is related to a parameter only via the measurement model
h( · ). Furthermore, only an important and special case where the measurement
model is linear in x is considered. The linear measurement model was given
in (2.23b) and is repeated here for convenience
yk = Hk (θ)xk + ek . (3.4)
Another example, where both the ls and the wls estimators are applied, is given
in Paper C. The left and right borders of a road are modeled by polynomials and
the coefficients are the parameters which are estimated given a batch of measure-
ments from a radar.
Put into words, the estimate is chosen to be the parameter most likely to produce
the obtained measurements.
The posterior p(xk |y1:k ), when xk is random, contains all known information
about the state of the target at time k. The maximum a posteriori (map) esti-
mator is defined by
x̂MAP
k = arg max p(xk |y1:k ) = arg max p(y1:k |xk )p(xk ), (3.12)
xk xk
or put in words, find the most likely estimate of the parameter given the measure-
ments y1:k . Bayes’ theorem (3.2) and the fact that the maximization is performed
over xk is used in the second equality of (3.12). The ml estimate is for instance
used in Paper E.
↑ ↑
If p(yk |xk ), p(xk+1 |xk ) and p(xk ) are Gaussian and their corresponding process
and senor models are linear, as in (2.23), then (3.13a) and (3.13b) reduce to the
Kalman filter prediction and measurement update, respectively. The Kalman
filter is treated in Section 3.2.1. In contrast, if p(yk |xk ), p(xk+1 |xk ) and p(xk )
can be approximated by a Gaussian and their corresponding process and sensor
models are nonlinear (2.1), several approximations of (3.13a) and (3.13b) exist.
The two most common filters are the extended Kalman Filter and the unscented
Kalman filter, which are outlined in Sections 3.2.2 and 3.2.3, respectively. Other
methods, including methods that approximate other density functions than Gaus-
sian, are neatly covered by Hendeby (2008) and Schön (2006). The most pop-
ular approaches are the particle filter, which is covered in Section 3.2.4, and
the marginalized particle filter, see e.g., Arulampalam et al. (2002); Cappe et al.
(2007); Djuric et al. (2003); Karlsson (2005); Schön et al. (2005).
The linear state space representation subject to Gaussian noise, which was given
in (2.23), is the simplest special case when it comes to state estimation. The model
is repeated here for convenience;
xk+1 = Fk (θ)xk + Gku (θ)uk + Gkw wk , wk ∼ N (0, Qk ), (3.15a)
yk = Hk (θ)xk + Hku (θ)uk + ek , ek ∼ N (0, Rk ). (3.15b)
The linear model (3.15) has two important properties. All density functions in-
volved in the model and state estimation are Gaussian and a Gaussian density
function is completely parametrized by the mean and the covariance, i.e. the first
and second order moment. Hence, the Bayesian recursion (3.13) is simplified
into only propagating the mean and covariance of the involved probability den-
3.2 Filter Theory 37
↑ ↑
The linearization used in the ekf assumes that all second and higher order terms
in the Taylor expansion are negligible. This is certainly true for many systems,
but for some systems this assumption can significantly degrade the estimation
performance. Higher order ekf are discussed by Roth and Gustafsson (2011); Bar-
Shalom and Fortmann (1988); Gustafsson (2000). This problem will be revisited
in the next section.
If the noise is additive, then the noise covariances can be added directly to the
estimated covariances of the non-augmented sigma points.
There exist many possibilities to choose the sigma points, a thorough discussion
about different alternatives is presented by Julier and Uhlmann (2004). In the
present work only the standard form is reproduced. The basic principle is to
choose one sigma point in the mean of xa and 2na points symmetrically on a
given contour, described by the state covariance P a . The sigma points χ i and the
associated weights w(i) are chosen as
χ(0) = x̂a w(0) = w(0) (3.24a)
1 − w(0)
r !
na
χ(i) = χ(0) + Pa w(i) = (3.24b)
1 − w(0) i 2na
1 − w(0)
r !
(i+na ) (0) na
χ =χ − Pa w(i+na ) = (3.24c)
1 − w(0) i 2na
√
for i = 1, . . . , na , where ( A)i is the i th column of any matrix B, such that A = BBT .
The augmented state vector makes it possible to propagate and estimate nonlin-
40 3 Estimation Theory
ear influences that the process noise and the measurement noise have on the state
vector and the measurement vector, respectively. The weight on the mean w(0) is
used for tuning and according to Julier and Uhlmann (2004) preferable proper-
ties for Gaussian density functions are obtained by choosing w(0) = 1 − n3a . After
the sigma points have been acquired, the augmented state vector can be parti-
tioned according to
x
χk|k
a
χk|k = χkw . (3.24d)
e
χk+1
The rest of the ukf is summarized in Algorithm 5.
An advantage of the ukf, compared to the ekf, is that the second order bias cor-
rection term is implicitly incorporated in the mean estimate. Example 3.5 shows
an important problem where the second order term should not be neglected.
3.2 Filter Theory 41
Let the sensor be located at the origin and the target at (x, y) = (0, 1) in this simple,
and commonly used example (Julier and Uhlmann, 2004). Measurements may be
simulated by adding Gaussian noise to the actual polar value (r, ψ) = (1, π/2)
of the target localization. A plot of several hundred state estimates, produced
in a Monte Carlo simulation, forms a banana shaped arc around the true value
(x, y) = (0, 1), as shown in Figure 3.1. The azimuth error causes this band of
Cartesian points to be stretched around the circumference of a circle, with the
result that the mean of these points lies somewhat closer to the origin than the
point (0, 1). In the figure it is clearly shown that that the ut estimate (×) lies
close to the mean of the measurements (◦). Furthermore, it is shown that the
linearized state estimate (+) produced by the ekf is biased and the variance in
the y component is underestimated.
As a result of the linearization in the ekf, the second order terms are neglected,
which produces a bias error in the mean as shown in Example 3.5. In Julier and
Uhlmann (2004) it is shown how the ut in some cases calculates the projected
mean and covariance correctly to the second order terms.
0.8
0.6
y
0.4
0.2
0 Sensor
−0.2
−0.6 −0.4 −0.2 0 0.2 0.4 0.6
x
(a)
1.02
0.98
0.96
y
0.94
0.92
0.9
−0.3 −0.2 −0.1 0 0.1 0.2 0.3
x
(b)
This chapter begins by putting the filter, introduced in the last chapter about
estimation theory, into the target tracking framework. In Section 4.1 the focus
is on the filter and therefore only the single target case is considered. An au-
tomotive target tracking system, must obviously be able to handle multiple tar-
gets, because potentially more than one vehicle are surrounding the own vehicle.
Therefore the chapter is continued by discussing the extension to multi target
tracking in Section 4.2. When reading the first two sections, compare with the
block diagram of the target tracking framework illustrated in the introduction in
Figure 1.4. Finally, in Section 4.3 the system is further enlarged to also be able to
track the shape and size of the the targets. This research area is called extended
target tracking.
45
46 4 Target Tracking
observation space
zk+1
zk
state space
target motion
xk+1
xk
Figure 4.1: Single target tracking problem. The upper layer is the observa-
tion space with the measurements and the lower layer is the state space with
the state variables.
The single target Bayes filter recursion consists of two steps, the prediction step
and the correction step. Since all variables are assumed to be unknown stochastic
variables in the Bayesian framework it is necessary to describe a priori informa-
tion about the state in the form of a prior density p0 (x0 ). The recursion was
illustrated with a flow chart in Section 3.2. A special case is the Kalman filter,
where the densities are assumed to be Gaussian.
4.2 Extension to Multitarget Tracking 47
it must also estimate the number of targets Nx , and find likely associations of the
measurements to the tracks. The latter uncertainty is known as the data associa-
tion problem.
A typical multi target system is shown in Figure 4.2, where several targets exist
at each time step. Aspects to consider in a multitarget tracking filter are listed
below:
• The number of targets Nx changes over time, for instance there are probably
more vehicles on the roads during rush hour than in the night.
• New targets appear and old targets disappear, as they enter or leave the
field of view of the sensor. Compare with Figure 4.2, where target x(4) ap-
pears at time k + 1.
• The system fails to detect targets because they are temporarily occluded or
because the sensor data is very noisy.
• The sensors receive a set of spurious measurements, also denoted clutter,
stemming from false reflections or from other type of objects in the field
of view, etc. In Figure 4.2 there are more observations than targets at each
time step.
In this section first the data association problem is treated in Section 4.2.1 and
thereafter an overview of the track management, which takes care of the prob-
lems in the list above is given in Section 4.2.2.
Figure 4.2: Multi target tracking system, several state vectors are present at
each time step and producing more than one observation. Note that there
are more observations than targets. A new target x(4) appears at time k + 1.
where γG is the gating threshold or gate size. Given a certain probability that a
true measurement produced by a target i will fall inside its gate, and that the
2
assumption di,j ∼ χn2y holds, a suitable gate threshold can be determined. The
(j) (i)
measurements zk ∈ Gi are considered as candidates for updating the track xk in
the data association algorithm. In other words the gating is a hard decision about
which measurements are feasible measurements for a target.
Now, different conflicts occur. There are several measurements falling within the
4.2 Extension to Multitarget Tracking 49
same gate and there are also measurements falling within more than one gate.
There exist many techniques to solve these conflicts, which are considered to be
the main part of the data association process. The simplest association algorithm
is called nearest neighbor (nn). This approach searches for a unique pairing,
(i) (j)
i.e., one track xk is only updated by at most one observation zk . There are
some possibilities to decide which measurement actually is the nearest. Common
(i,j)
approaches are to choose the measurement with the smallest error z̃k|k−1 or the
(i,j)
smallest statistical distance d 2 (z̃k|k−1 ), defined in (4.5), which is also known as the
Mahalanobis distance, see e.g., Bar-Shalom et al. (2001). The association distance
for each individual track is locally minimized separately with the nn approach,
which could lead to that two tracks are associated with the same measurement.
To avoid this problem it would be more beneficial to find the global minimum
distance considering all tracks simultaneously, and restrict that a measurement
can only be associated with one track. This approach is referred to as global
nearest neighbor (gnn), and its aim is to solve
Nk
X
2
min = di,λ i
(4.7)
λ
i
There exist a further group of association methods, which uses all measurements
that fall inside the gate. In these methods each measurement j is weighted in
accordance with the probability that it originates from track i. The two most
well known algorithms are probabilistic data association (pda) (Bar-Shalom and
Tse, 1975) and joint probabilistic data association (jpda) (Fortmann et al., 1983),
50 4 Target Tracking
(4) (4)
zk (6)
zk (6)
zk zk
(3) (3)
zk
(1) (1 ) zk (1)
zk (1 ) zk
ẑk |k −1 (2)
ẑk |k −1 (2)
ẑk |k −1 ẑk |k −1
(5) (5)
(2)
zk (2)
zk
zk zk
Figure 4.3: An example with two targets (blue and green) and six measure-
ments (in red). The ellipses are the gates and the two figures show the differ-
ence between the association methods nn and gnn.
where the difference lies in that pda considers each track separately, whereas
jpda forms global hypotheses to calculate the probabilities. Note that the number
of targets is assumed fixed in the above mentioned association methods. These
methods are not further discussed in this thesis, hence a detailed description is
omitted at this place.
pear from a certain place, but in most cases it is not easy to know beforehand.
A practical and straightforward method is to use the first measurement as the
mean value of the prior distribution. Other statistics of the prior distribution
are still considered as design variables. A better approach may be considered in
the Kalman filter, where the prior distribution is Gaussian and may be chosen as
x0 ∼ N (0, P0 ). If P0 is chosen very large in comparison with the measurement un-
certainty, the first updated state estimate will largely be based on the associated
measurement.
In order to know if the first measurement stems from a true target or from clut-
ter, the track management system must observe the tentative track for a number
of samples. There exist different methods to decide if a tentative track shall be
transformed into a confirmed track. One of the simplest and most well known
is the so called M/N logic, where a track is confirmed if M measurements out of
N possible are associated with the track. Other common methods are based on
scoring measurements, with for example sequential probability ratio test (sprt),
which was first proposed by Wald (1945). The method is based on comparing
the ratio between the likelihood of the hypothesis that a track describes a true
target and the likelihood of the hypothesis that it is a false alarm. Furthermore,
there exist combined validation and association methods. Integrated probabilis-
tic data association (ipda) expresses both the probability of target existence and
data association based on pda.
The track deletion methods are similar to the validation methods, both the M/N
logic and the score based approaches, mentioned above, can be applied. More
information and details about track management can be found in various books
about target tracking, see e.g., Blackman and Popoli (1999).
4.2 Definition (Extended Target). A target which potentially gives rise to more
than one measurement per time step irrespective of whether the target’s extent is
explicitly modeled (and/or estimated) or not is referred to as an extended target.
The methods used to track extended target are very similar to the ones used for
tracking a group of targets moving in formation. Extended target tracking and
group tracking are thoroughly described in e.g., Ristic et al. (2004). The bibli-
ography of Waxman and Drummond (2004) provides a comprehensive overview
of existing literature in the area of group and cluster tracking. There exist some
different approaches to represent, i.e., to model, the extended targets, of which
five methods are described in this section.
The uncertainty about the exact position of the point feature is modeled accord-
4.3 Extended Target Tracking 53
ing to
Np
Y
p(P W
|dTWW , ψT ) = N (p(i),W |RW T (ψT )p(i),T + dTWW , wp I2 ), (4.12)
i=1
which means that the uncertainty is assumed isotropic around the mean location
of the point and with known variance wp .
z N
At each time step a set of Nz measurements Z = {zi }i=1 is received and has to be as-
sociated to the states. Not all measurements arise from a point feature, some are
due to false detections (clutter). The association hypotheses are derived through
some data association algorithm. In Vermaak et al. (2005) a method is proposed
where the association hypotheses are included in the state vector and the output
of the tracking filter is a joint posterior density function of the state vector and the
association hypotheses. Furthermore, a multi-hypothesis likelihood is obtained
by marginalizing over all the association hypotheses. An alternative solution is
also proposed using a particle filter, where the unknown hypotheses are sampled
from a well designed proposal density function.
An automotive radar sensor model developed for simulation purposes is pro-
posed in Bühren and Yang (2006), where it is assumed that radar sensors often
receive measurements from specific reflection centers on a vehicle. These reflec-
tion centers can be tracked in a filter and valuable information regarding the
vehicle’s orientation can be extracted as shown by Gunnarsson et al. (2007). A
difficulty in solving the data association problem is the large number of associ-
ation hypotheses available. To reduce the complexity Gunnarsson et al. (2007)
proposes an approach where detections are associated with reflector groups. The
spatial Poisson distribution, discussed in the subsequent section, is considered to
be inappropriate, since the number of vehicle detections is assumed essentially
known and not adequately modeled by a Poisson process.
In paper F the point features are denoted measurement generating points (mgp),
and they are positioned on the surface of the target. In the referred publication,
the mgps are not considered having a fixed position on the surface, but they are
instead defined as a random finite set of points on the one-dimensional surface of
the target. The positions of the points on the surface are estimated in the target
tracking filter, but they are of less importance, since they only serve as a means
to estimate the position, shape and size of the entire target.
dent sample from the spatial distribution. The spatial model could be a bounded
distribution, such as a uniform pdf or an unbounded distribution, such as a Gaus-
sian. The Poisson assumption allows the problem, or more specifically the evalu-
ation of the likelihood, to be solved without association hypotheses. The spatial
distribution is preferable where the point source models are poor representations
of reality, that is in cases where the measurement generation is diffuse.
In Gilholm and Salmond (2005) two simple examples are given. One where the
principle axis of the extended target is aligned with the velocity vector, i.e., the
target is represented by a one dimensional uniform stick model. In the other
example, a Gaussian mixture model is assumed for the target. A Kalman filter
implementation with explicit constructions of assignment hypotheses is derived
from the likelihood in Gilholm and Salmond (2005), whereas in Gilholm et al.
(2005), a particle filter is applied directly given the likelihood which is repre-
sented by the Poisson spatial model of the stick. Hence, the need to construct
explicit measurement-target assignment hypotheses is avoided in Gilholm et al.
(2005). Swain and Clark (2010) proposes instead a standard measurement model
but represents instead the extended targets as a spacial cluster process.
Boers et al. (2006) presents a similar approach, but since raw data is considered,
no data association hypotheses are needed. The method to use raw data, i.e., con-
sider the measurements without applying a threshold, is referred to as track be-
fore detect. A one dimensional stick target is assumed also by Boers et al. (2006),
but unlike Gilholm and Salmond (2005), the target extent is assumed unknown.
The state vector is given by the stick’s center position and velocity as well as the
stick’s extension according to
h iT
x = x y ẋ ẏ l . (4.13)
The process model is a simple constant velocity model and the length l is modeled
as a random walk. The likelihood function is given by the probability distribution
Z
p(z|x) = p(z|x̃)p(x̃|x)d x̃, (4.14)
where the spatial extension is modeled by the pdf p(x̃|x) and x̃ is assumed to be
a point source from an extended target with center given by the state vector x.
Hence, a measurement is received from a source x̃ with likelihood p(z|x̃).
Parr (2003), where the sensor not only provides measurements of point observa-
tions, but rather range, bearing and down-range extent. The prime motivation of
the study is to aid track retention for closely spaced moving targets. Furthermore,
the state vector includes the position, velocity and the size of the ellipse. An ekf
is used in Salmond and Parr (2003), but it is concluded that the filter may di-
verge under certain conditions, since the relation between the down-range extent
measurement of the target and the position and velocity coordinates in the state
vector is highly nonlinear. The same problem is studied in Ristic and Salmond
(2004), where a ukf is implemented and tested. Even though the ukf shows bet-
ter performance it is concluded that neither the ekf nor the ukf are suitable for
this problem. The problem is further studied by Angelova and Mihaylova (2008),
where other filter techniques, based on Monte Carlo algorithms, are proposed. In
this paper the size of the ellipse takes values from a set of standard values, i.e.,
the algorithm estimates the type of object from a list, under the assumption that
typical target sizes are known.
In the previous chapter it was shown how single target tracking filters can be
extended and used to handle multiple targets by enclosing the filters with a tar-
get management framework. In this way measurements are associated to targets,
which are tracked as isolated entities. To be able to perform the data association
it is assumed that the number of present targets is known. This is rarely true, and
therefore a track management logic estimated the number of valid and confirmed
tracks. These methods explicitly separate the estimation of the number of targets
and estimation of target states, which is suboptimal. Random finite set (rfs)
formalism introduces a more rigorous approach to the multi target tracking prob-
lem. A set contains the random number of single target tracks, of which each is a
random vector. The state vectors are elements of the rfs and their change in time
is described with a motion model. The overall tracking problem is to compute
the posterior density of the set-valued quantity, which also enables the approach
to describe the uncertainty in the number of objects. The ordering of each single
target state is not of importance in the the set notation; the estimation problem is
reduced to capture the most essential part of the multi target tracking problem,
estimating the number of objects and their individual states.
This chapter is split into three sections, which takes the estimation problem from
a pure theoretical view into a practical implementation. The chapter begins in
Section 5.1 with an introduction to the Bayes formulation of the rfs filter, which
propagates the complete density of the rfs. This approach is not practically im-
plementable, and therefore a solution is to only propagate the first order mo-
ment, called the probability hypothesis density (phd), as described in Section 5.2.
There exist a few practical representations of the phd, of which the Gaussian mix-
ture has turned out to be the most well-used during the last years. The Gaussian
mixture phd filter (gm-phd) is the topic of Section 5.3. This chapter aims at sum-
57
58 5 PHD Filter and Its Implementation
observation space
Zk+1
Zk
state space
Figure 5.1: Illustration of the rfs of states and measurements at time k and
k + 1. Note that this is the same setup as previously shown for the standard
multitarget case in Figure 4.2.
marizing the basic ideas and implementation methods, more details can be found
in the textbook by Mahler (2007a), and other overviews are given by Mahler
(2009b); Challa et al. (2011). Point process theory is described by e.g., Daley
and Vere-Jones (2003); Streit (2010).
The conceptual idea of fisst is to redefine the target set as a single target, a so
called meta target, with multitarget state
X = {x(1) , x(2) , . . . , x(Nx ) }, (5.1)
a so called meta state; and similarly to redefine the observation set
Z = {z(1) , z(2) , . . . , z(Nz ) }, (5.2)
as a single measurement, a meta measurement, of the observed meta targets. The
concept is illustrated in Figure 5.1, where meta targets are shown in the state
space in the lower layer and meta observations in the observation space in the
upper layer. Furthermore, the multitarget multisensor data can be modeled using
a multisensor-multitarget measurement model
Zk = H(Xk ) ∪ Ck , (5.3)
5.1 Introduction to Finite Set Statistics 59
where H(Xk ) is the rfs of measurements originating from the true target and C
is the clutter rfs. To see the similarity, compare this multitarget model with the
single target measurement model (2.1a). The motion of the meta target can be
modeled using a multitarget motion model
Xk+1 = F(Xk ) ∪ Bk (5.4)
where F( · ) models the change of a rfs from time k to k + 1, and B is the rfs
of born targets. Compare, also here, with the single target motion model (2.1b).
Given this the multisensor, multitarget estimation problem can be reformulated
into a single-sensor, single-target problem. The rfs is formally defined in Sec-
tion 5.1.1, followed by a definition of the belief-mass function and multitarget
density function in Section 5.1.2. The multisource multitarget Bayes filter is de-
scribed in Section 5.1.3.
Since the definition may seem a bit abstract it is illustrated with a simple example,
related to target tracking.
5.2 Example: Target Tracking
Consider the target tracking example, which was introduced in Example 1.2. In
the example the state vector represents the Cartesian position of a target, i.e.,
nx = 2. The state space is defined as an Euclidean vector space, i.e., x ∈ R2 . This
state space is the underlying space Y0 in the definition above. The hyperspace
Y includes all finite subsets of Y0 = Rnx . Let the state variable vectors be x(i) ∈
Y0 = R2 for i = 1, . . . , ∞, then some realizations X of the random set X can be
X = ∅ (no targets), X = {x(1) , x(2) } (two targets with states x(1) and x(2) ), or X =
{x(1) , . . . , x(Nx ) } (Nx targets with states x(1) , . . . , x(Nx ) ).
and
δβΨ
pΨ (Y ) =(∅). (5.8)
δY
δ·
R
Here, · δY denotes the set integral, and δY denotes the set derivative. The
definitions of the set integral and the set derivative are too complicated to present
here, the interested reader can consult Mahler (2007a).
These definitions have the following practical use when applied to the likelihood
function and the Markov transition density
• In the single target case the probabilistic-mass function of the sensor model
p(S|x) = Pr(z ∈ S|x) is the probability that the random observation z will
be found in a given region S if the target state is x. Analogously, in the
multitarget case the belief-mass function β(S|X) = Pr(Z ⊆ S|X) is the total
probability that all observations will be found in any given region S, if the
multitarget state is X. In the single target case the likelihood function is
obtained by differentiation of the probabilistic-mass function. The same
concept holds for the multitarget case, i.e., the true multitarget likelihood
function p(Z|X) can, using fisst, be derived from β(S|X) according to
δβk|k (∅|Xk )
pk|k (Zk |Xk ) = (5.9)
δZk
δ
where δZk is the set derivative.
ing p(S|xk ) the Markov transition density pk+1|k (xk+1 |xk ) can be derived in
the single target case. Using fisst the multitarget Markov density can be
derived by taking the set derivative of the belief-mass function
δβk+1|k (∅|Xk )
pk+1|k (Xk+1 |Xk ) = (5.10)
δXk+1
Consider the single target state propagation summarized in Section 4.1. The filter
recursion is extended to the multitarget case under the fisst assumptions in this
section. The single target random state variables x are substituted with the rfs
of a set of targets. Before giving the filter recursion, the interpretation of a set
probability function is discussed. Consider the multitarget state propagation
predictor corrector
· · · → pk|k (Xk |Z1:k ) → pk+1|k (Xk+1 |Z1:k ) → pk+1|k+1 (Xk+1 |Z1:k+1 ) → · · ·
i.e., the set density is not only a function of the state variable values, but also a
function of the number of targets Nx in the set.
The single target Bayes filter equations were stated in (3.13), the multitarget
Bayes filter counterpart has the form
pk (Z1:k |Xk ) · pk|k−1 (Xk |Z1:k−1 )
pk|k (Xk |Z1:k ) = R (5.11a)
pk (Z1:k |Xk ) · pk|k−1 (Xk |Z1:k−1 ) δXk
Z
pk+1|k (Xk+1 |Z1:k ) = pk+1|k (Xk+1 |Xk ) · pk|k (Xk |Z1:k )δXk (5.11b)
where pk (Zk |Xk ) is the multisource likelihood function, pk+1|k (Xk+1 |Xk ) is the
multitarget Markov transition function, pk|k−1 (Xk |Zk−1 ) is the multitarget prior,
pk|k (Xk |Zk ) is the multitarget posterior and pk+1|k (Xk+1 |Zk ) is the multitarget pre-
62 5 PHD Filter and Its Implementation
↑ ↑
compare with the single target filter recursions given on the same form in Sec-
tion 3.2. Note that the motion model describes a union of the set of persisting
targets and the set of new born, or detected, targets. In the same manner, the
measurement model describes a union of the set of target generated observations
and the set of clutter measurement. Here, the measurement set Z is the set of all
observations received by the sensor.
The conclusions of this section can be summarized in a few items, which will be
furher addressed in the next section:
• The Bayesian filter on sets is very similar to standard Bayesian density re-
cursion.
• Related single target densities and operations are replaced with their set
equivalents, compare (5.11) with (3.13).
R
• A set integral · δX is necessary to derive the posterior
Z
p(Xk |Z1:k ) ∝ p(Zk |Xk ) p(Xk |Xk−1 )p(Xk−1 |Z1:k−1 )δXk−1
Based on the same analogy as with the single target tracking the density of the
rfs my be compressed to the first order moments, and these may be propagated
as follows:
predictor corrector
· · · →pk|k (Xk |Z1:k ) → pk+1|k (Xk+1 |Z1:k ) → pk+1|k+1 (Xk+1 |Z1:k+1 )→ · · ·
↓ ↓ ↓
predictor corrector
Dk|k (xk |Z1:k ) → Dk+1|k (xk+1 |Z1:k ) → Dk+1|k+1 (xk+1 |Z1:k+1 )
The first order moment is called a phd and denoted D. However, it is not obvious
what the multitarget counterpart of an expected value is; a naïve defininition
would be
Z
E(Ψ ) = X · pΨ (X)δX, (5.12)
where pΨ (X) is the multitarget probability density of a rfs Ψ . However, this inte-
gral is not mathematically defined, since addition of finite subsets is not usefully
defined. A strategy is to transform subsets X into vectors TX in some vector space.
It is important that this transformation X → TX preserves set-theoretic structures,
i.e., transforming unions into sums TX∪X 0 = TX + TX 0 , whenever X ∩ X 0 = ∅. An
indirect expected value can now be defined according to
Z
E(Ψ ) = TX · pΨ (X)δX. (5.13)
where δy (x) is the Dirac delta function with peak at y. Given these assumptions,
the multitarget analog of the expected value is given by
Z
DΨ (x) = δX (x) · pΨ (X)δX. (5.15)
The expected value is illustrated in Figure 5.2b. Note also that the expected num-
ber of targets in the volume S is
Z
bx = Dk|k (x)dx
N (5.18)
S
In fact one can define the phd using the expected number of Rtargets as follows:
Given any region S of single target state space X0 the integral S Dk|k (x)dx is the
expected number of targets in S.
The phd is an intensity function, which models the joint density over the states.
The phd is multi-modal, with peaks near the actual targets, see Figure 5.3. This
figure shows the phd approximation of the rfs in Figure 5.1. The intensity func-
tion is used in Paper D to represent the density of stationary targets along the
edges of a road. An autonomous driving vehicle should avoid areas with high
intensity and the objective should be to keep the vehicle in the valleys of the
intensity function.
The phd filter consists of two equations, the predictor equation, which extrapo-
lates the current phd to the predicted phd at the time of the new observation,
and the corrector equation, which updates the predicted phd with the new obser-
vations.
The phd predictor equations takes care of the motion of existing targets, birth
and spawn of new targets and the death of existing targets. The components of
the prediction model are summarized below:
5.2 Introduction to the PHD filter 65
10
0
1
1
0.5
0.5
0 0
(a) Sum of Dirac functions δXk
10
0
1
1
0.5
0.5
0 0
(b) Expectation E(δXk )
Figure 5.2: Definition and example of the phd, the Dirac functions repre-
sents the rfs and the surface the first order moment, i.e., the phd.
66 5 PHD Filter and Its Implementation
Figure 5.3: Illustration of the phd for the time steps k and k + 1. This is the
pdf of the rfs in Figure 5.1.
In the table above PS (x) denotes the probability that a target with state x at time k
will survive at time k + 1, p(xk+1 |xk ) is the single target Markov transition density,
b(Xk+1 |xk ) is the phd of targets spawned by other targets with state x, and b(Xk+1 )
is the phd of new appearing targets. The phd predictor equation is given by
Dk+1|k (xk+1 |z1:k ) =
| {z }
time updated PHD
Z h i
= bk+1|k (xk+1 ) + PS (ξk ) p(xk+1 |ξk ) + b(xk+1 |ξk ) Dk|k (ξk |z1:k ) dξk .
| {z } | {z } | {z } | {z } | {z }
target birth PHD probability of Markov transition target spawn by PHD from previous k
survival density existing targets
(5.19)
The parts of the model are illustrated with a number of plots in Figure 5.4. One
example of a phd Dk|k (ξ|z1:k ) from the previous time step k is illustrated in Fig-
ure 5.4a. The magnitude of the phd is reduced a bit when multiplied with the
probability of survival PS (ξ) in Figure 5.4b. Thereafter, multiplying with the
Markov transition density and then integrating changes the mean value and the
covariance, see Figure 5.4c. Note, that all target motions are assumed statistically
independent. In Figure 5.4d, the spawn phd b(xk+1 |ξk ) is added in red, and in
Figure 5.4e the birth phd bk+1|k (xk+1 ) in green. Finally, the total predicted phd
Dk+1|k (xk+1 |z1:k ), after time update, is shown in Figure 5.5a.
5.2 Introduction to the PHD filter 67
0.5 0.5
0.4 0.4
PD Dk−1
0.3 0.3
Dk−1
0.2 0.2
0.1 0.1
0 0
−5 0 5 10 −5 0 5 10
x x
0.5 0.5
∫ (PD p(xk|ξ)+b(xk|ξ) Dk−1 d ξ
0.4 0.4
∫ PD p(xk|ξ)Dk−1 d ξ
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−5 0 5 10 −5 0 5 10
x x
0.5 0.5
∫ (PD p(xk|ξ)+b(xk|ξ) Dk−1 d ξ + b
0.4 0.4
0.3 0.3
Dk|k−1
0.2 0.2
0.1 0.1
0 0
−5 0 5 10 −5 0 5 10
x x
The phd corrector equations takes care of the likelihood of the existing targets,
misdetection of targets and clutter. The components of the prediction model are
summarized below:
model update description
likelihood: p(zk |xk ) xk → zk likelihood that target will gener-
ate observation zk if it has state
xk
misdetection: (1 − PD (xk )) xk → ∅ state dependent probability that
target will not generate an obser-
vation
clutter: c(Zk ) ∅ → Zk likelihood that a set Zk =
(1) (M)
{zk , . . . , zk } of clutter observa-
tions will be generated
In the table above PD (xk ) is the probability of detection of a target with the state
x at time step k, p(zk |xk ) is the single target likelihood function and c(Zk ) is the
density of Poisson false alarms, due to clutter. The modeling is done under the
assumption that the observations and the clutter are statistically independent.
Furthermore it is assumed that the multitarget prediction is approximately Pois-
(1) (M)
son. Given a new scan of data Zk = {zk , . . . , zk } the corrector equation is given
by
X Λk|k (x|z)
Dk|k (x|Z1:k ) ≈ R + (1 − PD (x)) Dk|k−1 (x|Z1:k−1 ) (5.20)
| {z } z∈Zk λk ck (z) + Λk|k (x|z)dx | {z }
updated PHD | {z } predicted PHD
pseudo-likelihood
0.5 0.5
0.4 0.4
(1−PD)Dk|k−1
0.3 0.3
Dk|k−1
0.2 0.2
0.1 0.1
0 0
−5 0 5 10 −5 0 5 10
x x
0.5 0.5
0.4 0.4
Σ Dk|k(x|z)
0.3
p(z|x)
0.3
0.2 0.2
0.1 0.1
0 0
−5 0 5 10 −5 0 5 10
x x
(c) likelihood functions of two measure- (d) Detected and non-detected targets
ments
0.5
0.4
0.3
Dk
0.2
0.1
0
−5 0 5 10
x
Figure 5.5: Steps in the phd update and corrector equations (5.20).
70 5 PHD Filter and Its Implementation
the present chapter, and it is also not used in any of the publications in Part II.
Details about the cphd filter can be found in Mahler (2007b,a), and a Gaussian
mixture implementation is described in Vo et al. (2007).
The standard phd filter as described in this section does not maintain track labels
from time step to time step. A so called peak to track association technique has
been proposed by Lin et al. (2004); Panta et al. (2004) in order to maintain track
labels.
Furthermore, phd smoother has been studied in Mahler et al. (2010); Clark (2010),
and a Gaussian mixture implementation is presented by Vo et al. (2010, 2011).
Again, since the smoother is not further used in Part II, it is also not described
here.
The so called intensity filter, developed by Streit, is another generalization of the
phd filter. The difference is that in the phd filter certain prior assumptions re-
garding target birth and measurement clutter are made. More specifically the
intensity filter uses an augmented single target state space, while the phd filter
uses only the standard single target state space. The augmented state space rep-
resents the absent target hypothesis, and it therefore allows the filter to on-line
estimate the intensities of the target birth and measurement clutter Poisson point
processes. Note that the phd is an intensity function, and in the phd filter the
birth and clutter intensities are known a priori. A thorough description of the
intensity filter is given in the textbook by Streit (2010) and a summary is given in
the publications Streit and Stone (2008); Streit (2008).
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
Dk
Dk
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−5 −4 −3 −2 −1 0 1 2 3 4 5 −5 −4 −3 −2 −1 0 1 2 3 4 5
state space state space
(a) Sequential Monte Carlo phd approx- (b) Gaussian mixture phd approxima-
imations (smc-phd) tion (gm-phd)
↓ ↓ ↓
PJk|k (i)
(i) (i)
PJk+1|k (i)
(i) (i)
PJk+1|k+1 (i)
(i) (i)
··· → w N x; µ , P → w N x; µ ,P → w N x; µ ,P
i=1 k|k k|k k|k i=1 k+1|k k+1|k k+1|k i=1 k+1|k+1 k+1|k+1 k+1|k+1
↓ ↓ ↓
The phd is represented by a Gaussian mixture and the summary statistics of the
Gaussian components, i.e., the mean value µ and the covariance P are propagated
together with a weight w for each Gaussian component. The number of Gaussian
components at each time step is denoted J. The prior, predicted and posterior
gm-phd are shown above.
When deriving the gm-phd filter some assumptions are made, see Vo and Ma
(2006). Since these assumptions are formally repeated in Paper E, on Page 230,
72 5 PHD Filter and Its Implementation
these are only summarized here. It is assumed that each target follows a linear
Gaussian motion model,
pk+1|k (xk+1 |xk ) = N (xk+1 ; Fk xk , Qk ), (5.22)
where Fk is the state transition matrix and Qk is the process noise covariance, and
it is assumed that all target motions are statistically independent. Observations
and clutter are assumed statistically independent and the sensor is assumed to
have a linear Gaussian measurement model, i.e.,
pk|k (zk |xk ) = N (zk ; Hk xk , Rk ), (5.23)
where Hk is the measurement model matrix and Rk is the observation noise co-
variance. Furthermore it is assumed that the survival and detection probabilities
are state independent, i.e., PS (x) = PS and PD (x) = PD . The phd of the birth and
the spawn are Gaussian mixtures
Jb,k
(i) (i) (i)
X
b(Xk+1 ) = wb,k N (x; µb,k , Pb,k ), (5.24)
i=1
Jβ,k+1
(i) (`) (i,`) (i,`)
X
b(Xk+1 |xk ) = wβ,k+1 N (x; µβ,k+1|k , Pβ,k+1|k ), (5.25)
`=1
(i) (i) (i)
where wb,k , µb,k and Pb,k are the weight, the mean and the covariance of the ith
born component, i = 1, . . . , Jb,k , and Jb,k is the number of components. Further,
(`) (i,`) (i,`)
wβ,k+1 , µβ,k+1|k and Pβ,k+1|k are the weight, mean and covariance of the `th com-
ponent ` = 1, . . . , Jβ,k+1 at time k + 1 spawned from the ith component at time k,
and Jβ,k+1 is the total number of components spawned from component i. It is
also assumed that the predicted multitarget rfs are Poisson.
In this section the gm-phd filter algorithm including the predictor and the cor-
rector equations, are given. The prediction equation is an approximation of the
general phd prediction equation (5.19), with the difference that the phd compo-
nents are substituted with Gaussian mixtures according to
Jb,k Jk−1
(j) (j) (j) (j) (j) (j)
X X
Dk|k−1 (x|Z1:k−1 ) = wb,k N (x; µb,k , Pb,k ) + PS wk−1 N (x; µk|k−1 , Pk|k−1 )
j=1 j=1
Jβ,k
Jk−1 X
(j,`) (j,`) (j,`)
X
+ wβ,k|k−1 N (x; µβ,k|k−1 , Pβ,k|k−1 ). (5.26)
j=1 `=1
In the case when the motion and sensor models are linear, the Gaussian compo-
nents can be predicted using the prediction step of the Kalman filter (3.16). The
illustration in Figure 5.4 still holds for this Gaussian case, and the discussion is
therefore not repeated again. The corrector equation is an approximation of the
5.3 Gaussian Mixture Implementation 73
general phd corrector equation (5.20), with the difference that the phd compo-
nents are substituted with Gaussian mixtures according to
JX
k|k−1 XNz JX
k|k−1
(i) (i) (i) (i,j) (i,j) (i,j)
Dk|k (x|Z1:k ) ≈ wk|k N µk|k , Pk|k + wk|k N µk|k , Pk|k (5.27)
i=1 j=1 i=1
| {z } | {z }
not detected targets detected targets
To allow for nonlinear motion and sensor models, the ekf or ukf can be utilized,
in that case the prediction and update of the single Gaussian components are in-
stead performed as described in Algorithm 4 and 5, respectively. The ukf version
of the gm-phd filter is used in Paper D to track the edges of a road and thereby
to create an intensity map of the environment. An extension of the kf based
gm-phd used to track extended targets, is presented Paper E. In that paper mod-
ifications are made to the update step to account for the extended targets, which
may give rise to more than one measurement per target and time step. In Paper F
the gm-phd filter is used as an inner core in a tracking filter which aims at esti-
mating the size and shape of targets. The outer shell of the approach represents
the position and motion of the target, whereas the phd is defined on the surface
of the target as a means to estimate its shape.
than some threshold, see Algorithm 9. An advantage with the phd filter is that
a hard decision, i.e., extraction of targets, is only performed for obtaining an out-
put from the filter at each time step. The extracted targets are not used within the
filter; the information about the intensity of targets remains in the filter without
the need to make hard decisions, which would destroy information.
76 5 PHD Filter and Its Implementation
6.1 Conclusion
The work presented in this thesis has dealt with the problem of estimating the
motion and parameters of a vehicle, as well as representing and estimating its
surroundings. In this context the surroundings consist of other vehicles, station-
ary objects, and the road. Here, a major part of the work is not only the estima-
tion problem itself, but also the way in which to represent the environment, i.e.,
building maps of stationary objects and describing the size of other vehicles.
The second part of the thesis begins in Paper A with an overview of mapping
techniques used to describe a road and its closer environment. Four different
types of mapping philosophies are described; beginning with so called feature
based maps, where each element of the map both described the properties and
location of the specific map-element. In location based maps each map-element
already corresponds to a certain location and the value of the element describes
only the properties in that position. Hence, feature based maps are usually rather
sparse. Furthermore, road maps and intensity based maps are also summarized.
The intensity based map is the main topic of paper D.
Paper B is concerned with estimating the lane geometry. The lane markings are
described by a polynomial and the coefficients are the states to estimate. This
77
78 6 Concluding Remarks
problem can be solved with a camera and computer vision, but by fusing the
data obtained from the image processing with information about the ego vehi-
cle’s motion and the other vehicles’ movement on the road, the road geometry
estimate can be improved. The other vehicles are tracked primarily by using
measurements from a radar. The motion of the ego vehicle is estimated by com-
bining measurements from the vehicle’s imu, steering wheel angle sensor and
wheel velocity sensors in a model based filter. The model is in this case the so
called single track model or bicycle model.
Paper C and D both deal with the problems of estimating and representing sta-
tionary objects along the edges of a road. Paper C, associates the radar measure-
ments to extended stationary objects in the form of curved lines and tracks these
lines as extended targets. The lines are modeled as polynomials and one novelty
in this paper is that the approach leads to a so called errors in variables problem
when noisy point measurements are associated to polynomials. The standard ap-
proach is to assume that the measurements are only noisy in one direction when
associated to a line, here however noise is assumed to be present in all directions.
This assumption is of course more realistic and the results are more accurate.
Paper D describes a new type of map, the intensity based map. In this paper the
environment, i.e., the stationary objects, is represented by an intensity function
or probability hypothesis density (phd) as it is also called. The intensity function
is described by a Gaussian mixture. This makes the representation very sparse,
since only a number of sufficient statistics must be tracked, but also very rich
since the extension of the Gaussian components can cover and model larger areas.
It is shown how prior knowledge of the road construction can be used to both
improve the update of the filter and to simplify the map representation.
Paper E and F deal with the tracking of extended moving targets. In Paper E the
target is still modeled as a point target, however, it allows for obtaining more than
one measurement per target. A modification of the pseudo-likelihood function
used in the update step of the so called phd filter is presented. In the standard
filter one measurement can only be obtained from each target per time step. The
novelty of the paper is the practical implementation of the pseudo-likelihood
function and partitioning of the measurement data. In a number of simulations
and experiments the advantages of the modifications are shown. A drawback is
that the wrong number of targets is estimated in some rare cases, this is further
discussed in the next section outlining some possible directions of future work.
In Paper F, the size and the shape of an extended target are also estimated. It
is still assumed that the sensor obtains a number of point measurements. In or-
der to associate the point measurements to the shaped target it is assumed that
the measurements have been produced by a number of measurement generating
points on the surface of the target. The measurement generating points are repre-
sented by a random finite set and these are only used as a means to estimate the
size and the shape of the target. Only one simple simulation example is shown
and also here future work will cover more situations.
6.2 Future Research 79
The last paper, Paper G, aims at estimating the wheel radii of a vehicle. This is
made under the assumption that a change in the wheel radii is related to a change
in tire pressure, and that tire pressure losses can be detected using this approach.
The proposed method is only based on measurements from the wheel speed sen-
sors and the gps position information. The wheel radii are represented with noise
parameters in the state space model. The novelty lies in a Bayesian approach to
estimate these time-varying noise parameters on-line using a marginalized parti-
cle filter. Experimental results of the proposed approach show many advantages,
both regarding the accuracy and the reduced computational complexity, when it
is compared with other methods.
The approaches in Papers A-D, and G have been evaluated on real data from both
freeways and rural roads in Sweden. In Paper E data of moving pedestrians has
been used. Of the seven publications in Part II six are journal papers and one is
a peer-reviewed conference paper. Five of the papers are published and two are
still under review.
the ones used in this thesis. A very interesting avenue for future work is to com-
bine the ideas presented in this thesis with information from a camera about the
height differences on the road side within a sensor fusion framework. This would
probably improve the estimates, especially in situations when there are too few
radar measurements available.
Bibliography
A. Abdulle and G. Wanner. 200 years of least squares method. Elemente der
Mathematik, 57:45–60, May 2002.
D. Angelova and L. Mihaylova. Extended object tracking using Monte Carlo meth-
ods. IEEE Transactions on Signal Processing, 56(2):825–832, February 2008.
81
82 Bibliography
F. Bengtsson and L. Danielsson. Designing a real time sensor data fusion system
with application to automotive safety. In Proceedings of the World Congress
on Intelligent Transportation Systems and Services, New York, USA, November
2008.
D. P. Bertsekas. The auction algorithm for assignment and other network flow
problem: a tutorial. Interfaces, 20(4):133–149, July 1990.
M. Bühren and B. Yang. Simulation of automotive radar target lists using a novel
approach of object representation. In Proceedings of the IEEE Intelligent Vehi-
cles Symposium, pages 314–319, Tokyo, Japan, 2006.
S. J. Julier and J. K. Uhlmann. New extension of the Kalman filter to nonlinear sys-
tems. In Signal Processing, Sensor Fusion, and Target Recognition VI, volume
3068, pages 182–193, Orlando, FL, USA, 1997. SPIE.
S. J. Julier and J. K. Uhlmann. Reduced sigma point filters for the propagation of
means and covariances through nonlinear transformations. In Proceedings of
the American Control Conference, volume 2, pages 887–892, 2002.
T. Kailath. Linear systems. Prentice Hall, Englewood Cliffs, NJ, USA, 1980.
L. Ljung. System identification, Theory for the user. System sciences series. Pren-
tice Hall, Upper Saddle River, NJ, USA, 2 edition, 1999.
Bibliography 87
C. Lundquist and R. Großheim. Method and device for determining steering an-
gle information. German Patent Application DE 10 2007 000 958 Mai 14, 2009,
International Patent Application WO 2009 047 020 April 16, 2009 and Euro-
pean Patent Application EP 2205478 April 16, 2009, Priority date October 2,
2007.
C. Lundquist and W. Reinelt. Back driving assistant for passenger cars with
trailer. In Proceedings of the SAE World Congress, SAE paper 2006-01-0940,
Detroit, MI, USA, April 2006a.
R. P. S. Mahler, B.-N. Vo, and B.-T. Vo. The forward-backward probability hy-
pothesis density smoother. In Proceedings of the International Conference on
Information Fusion, pages 1–8, Edinburgh, UK, July 2010.
S. Malinen, C. Lundquist, and W. Reinelt. Fault detection of a steering wheel
sensor signal in an active front steering system. In Preprints of the IFAC Sym-
posium on SAFEPROCESS, pages 547–552, Beijing, China, August 2006.
M. Mitschke and H. Wallentowitz. Dynamik der Kraftfahrzeuge. Springer, Berlin,
Heidelberg, 4 edition, 2004.
E. Nilsson, C. Lundquist, T. B. Schön, D. Forslund, and J. Roll. Vehicle motion
estimation using an infrared camera. In Proceedings of the World Congress of
the International Federation of Automatic Control, Milan, Italy, August 2011.
U. Orguner, C. Lundquist, and K. Granström. Extended target tracking with a
cardinalized probability hypothesis density filter. In Proceedings of the In-
ternational Conference on Information Fusion, pages 65–72, Chicago, IL, USA,
July 2011.
E. Özkan, C. Lundquist, and F. Gustafsson. A Bayesian approach to jointly esti-
mate tire radii and vehicle trajectory. In Proceedings of the IEEE Conference
on Intelligent Transportation Systems, Washington DC, USA, October 2011.
K. Panta, B.-N. Vo, S. Singh, and A. Doucet. Probability hypothesis density filter
versus multiple hypothesis tracking. In Signal Processing, Sensor Fusion, and
Target Recognition XIII, volume 5429, pages 284–295, Orlando, FL, USA, 2004.
SPIE.
R. Rajamani. Vehicle Dynamics and Control. Springer, Boston, MA, USA, 2006.
ISBN 978-0-387-28823-9.
G. Reimann and C. Lundquist. Method for operating electronically controlled
servo steering system of motor vehicle, involves determining steering wheel
angle as measure for desired steering handle angle by steering handle for steer-
ing wheels of motor vehicle. German Patent Application DE 102006053029
May 15, 2008, Priority date November 10, 2006.
W. Reinelt and C. Lundquist. Observer based sensor monitoring in an active front
steering system using explicit sensor failure modeling. In Proceedings of the
IFAC World Congress, Prague, Czech Republic, July 2005.
W. Reinelt and C. Lundquist. Mechatronische Lenksysteme: Modellbildung und
Funktionalität des Active Front Steering. In R. Isermann, editor, Fahrdy-
namik Regelung - Modellbildung, Fahrassistenzsysteme, Mechatronik, pages
213–236. Vieweg Verlag, September 2006a.
W. Reinelt and C. Lundquist. Controllability of active steering system hazards:
From standards to driving tests. In Juan R. Pimintel, editor, Safety Critical
Automotive Systems, pages 173–178. SAE International, 400 Commonwealth
Drive, Warrendale, PA, USA, August 2006b.
90 Bibliography
W. Reinelt and C. Lundquist. Method for assisting the driver of a motor vehicle
with a trailer when reversing. German Patent DE 10 2006 002 294 February 24,
2011, European Patent Application EP 1810913 July 25, 2007 and Japanese
Patent Application JP 2007191143 August 2, 2007, Priority date January 18,
2006.
B.-N. Vo, B.-T. Vo, and R. P. S. Mahler. A closed form solution to the probability
hypothesis density smoother. In Proceedings of the International Conference
on Information Fusion, pages 1–8, Edinburgh, UK, July 2010.
B.-N. Vo, B.-T. Vo, and R. P. S. Mahler. Closed form solutions to forward-backward
smoothing. IEEE Transactions on Signal Processing, 2011. doi: 10.1109/TSP.
2011.2168519.
B.-T. Vo, B.-N. Vo, and A. Cantoni. Analytic implementations of the cardinalized
probability hypothesis density filter. IEEE Transactions on Signal Processing,
55(7):3553–3567, July 2007.
N. Wahlström, J. Callmer, and F. Gustafsson. Single target tracking using vector
magnetometers. In IEEE Conference on Acoustics, Speech and Signal Process-
ing, pages 4332–4335, Prague, Czech Republic, May 2011.
A. Wald. Sequential tests of statistical hypotheses. The Annals of Mathematical
Statistics, 16(2):117–186, June 1945.
M. J. Waxman and O. E. Drummond. A bibliography of cluster (group) tracking.
In Proceedings of Signal and Data Processing of Small Targets, volume 5428,
pages 551–560, Orlando, FL, USA, April 2004. SPIE.
M. Wieneke and S. J. Davey. Histogram PMHT with target extent estimates based
on random matrices. In Proceedings of the International Conference on Infor-
mation Fusion, pages 1–8, Chicago, IL, USA, July 2011.
J. Y. Wong. Theory Of Ground Vehicles. John Wiley & Sons, New York, USA, 3
edition, 2001.
T. Zajic and R. P. S. Mahler. Particle-systems implementation of the PHD
multitarget-tracking filter. In Signal Processing, Sensor Fusion, and Target
Recognition XII, volume 5096, pages 291–299, Orlando, FL, USA, April 2003.
SPIE.
Z. Zomotor and U. Franke. Sensor fusion for improved vision based lane recogni-
tion and object tracking with range-finders. In Proceedings of the IEEE Confer-
ence on Intelligent Transportation Systems, pages 595–600, Boston, MA, USA,
November 1997.