Path Planning and Tracking Control

for an Automatic Parking Assist System

Emese Sz adeczky-Kardoss and B alint Kiss
Dept. of Control Engineering and Information Technology
Budapest University of Technology and Economics, Hungary
[email protected], [email protected]
Abstract. This paper presents two major components of an automatic
parking assist system (APAS). The APAS maps the environment of the
vehicle and detects the existence of accessible parking place where the
vehicle can park into. The two most important tasks the APAS must
then realize are the design of a feasible path geometry and the tracking
of this reference in closed loop such that the longitudinal velocity of the
vehicle is generated by the driver and the controller determines the steer-
ing wheel angle which is realized by an Electric Steering Assist System
(EPAS). We present in details the solution of the continuous curvature
path planning problem and the time-scaling based tracking controller.
These components are part of an implemented APAS on a commercial
passenger car including an intelligent EPAS.
Keywords. automatic parking assist system, continuous curvature path
planning, tracking control, time-scaling
1 Introduction
An important goal of automatic vehicle control is to improve safety and drivers
comfort. The APASs provide this for parking maneuvers which are usually per-
formed at low velocities. The APAS collects rst information about the envi-
ronment of the vehicle (position of the obstacles) which are necessary to nd a
parking lot and to complete a safe parking maneuver [1].
One may distinguish fully or semi-automated APASs according to the drivers
involvement during the maneuver. In the fully automatic case the APAS inu-
ences both the steering angle and the longitudinal velocity of the car. In the case
of vehicles without automatic gear the semi-automatic APAS is the only avail-
able option. In this case the driver needs to generate the longitudinal velocity of
the car with an appropriate management of the pedals, while the APAS controls
the steering wheel.
There exists now several APASs on the market. The Aisin Seiki Co. Ltd. has
developed a semi-automatic parking system for Toyota [2] where a camera ob-
serves the environment of the car. The Evolve project resulted a fully automatic
APAS for a Volvo type vehicle based on ultrasonic sensor measurements [3]. The
Volkswagen Touran may be ordered with an option that also assists parking
maneuvers [4]. This solution also uses ultrasonic sonars for the semi-automatic
parking maneuvers. We do not address in this paper the map making and park-
ing lot detection procedures and the signal processing problems related to the
map making from distance measurement, and the position and orientation esti-
mation of the car. The rst task is realized in our setup using ultrasonic sensors
and the second problem is solved by the use of the angular velocity data of the
ABS sensors of the car.
To design an APAS, a mathematical model of the car has to be given. Several
vehicle models are presented in the literature including kinematic and dynamic
models. Let us suppose that the kinematic models such as the ones reported
in [57] describe in a satisfactory way the behavior of the vehicle at low velocities
where parking maneuvers are executed.
If the model of the vehicle is known, then the path planning method and the
tracking controller algorithm can be designed based on the motion equations.
To plan the motion one can use deterministic [8] or probabilistic methods [9, 10]
as well. To get a continuous curvature path special curves should be used [11].
Algorithms based on soft computing methods can also calculate the reference
path [12, 13].
In the case of the semi-automatic systems the design of the tracking con-
troller is more involved than in the fully automatic case. The tracking controller
in the fully automatic case may inuence the behavior of the car using two
inputs whereas one losses one input, namely the longitudinal velocity in the
semi-automated case where the car velocity is generated by the driver hence the
velocity prole during the execution of the parking maneuver may be consider-
ably dierent from the one used for the path planning. This problem is addressed
by [14] using time-scaling.
Our goal was to develop a parking assist system which can operate both in
fully and semi-automatic mode in three dierent parking situations (parking in
a lane, parking in a row, and diagonal parking [15]).
In this paper we present rst the components of the system (Sect. 2). The
kinematic vehicle model is described in Sect. 3 and the path planning and track-
ing controller algorithms are detailed in Sect. 4-6. Section 7 presents some results
based on tests on a real car and a short summary concludes the paper.
2 Components of the System
To ensure autonomous behavior, several tasks have to be solved: the system
should be able to detect obstacles in its environment; it has to measure or esti-
mate its position and orientation; the reference motion has to be planned, and
nally, this reference should be tracked as accurately as possible. These tasks are
performed by separate interconnected subsystems which are depicted in Fig. 1.
The ABS sensors of the vehicle can detect the displacement of the wheels
of the car. Based on these data, an estimator calculates the actual position and
orientation of the car in a xed word coordinate frame. This estimated state is
used by the mapping and controller modules. To draw a map, additional data are
also required about the environment. Ultrasonic sensors are used to measure the
Autonomous Parking System
Map
constuction
Path
planning
State
estimator
Controller
Vehicle
Distance
measurements
Reference
path Control signals
Parking
place
State of
vehicle
Measured data
Environment
User interface
State of
vehicle
Fig. 1. Components of the Automatic Parking Assist System
distances to the surrounding obstacles. Based on these distance measurements
a map can be created. One may then use simple algorithms to detect accessible
parking places (if any) on the map.
During the motion planning a reference path is calculated which connects
the initial and the desired nal congurations. In this planning phase some
constraints (e.g. the non-holonomic behavior described by the model, collision
avoidance, maximal values of the actuator signals) have to be taken into con-
sideration. Finally, the tracking control algorithm is used to track the reference
path.
3 Kinematic Vehicle Model
Both the path planning and tracking control algorithms use the kinematic model
of the vehicle. To calculate the geometry of the path a slightly extended model
is used in order to ensure the continuous curvature property.
As it is usual in the literature, the reference point of the car is the midpoint
of the rear axle denoted by R. The conguration of the vehicle (q) is described
by four state variables: position of the reference point R (x, y), the orientation
of the car , and the curvature (), which is the inverse of the turning radius.
Supposing that the Ackermann steering assumptions hold true, the motion of
the vehicle can be described by the kinematic model of the bicycle tted on the
longitudinal symmetry axis of the car (see Fig. 2):
q =

x
y

cos
sin

v +

0
0
0
1

. (1)
The longitudinal velocity of the car is denoted by v and the velocity of the
change in curvature is given by . Let us denote the axle space by b. Then we
get the following relationship between the curvature and its derivative:
R
1
k
d
b
y
x
y
x
y
d
Fig. 2. Kinematic model of the vehicle in the x y plane
=
tan
b
, = =

b cos
2

, (2)
where is the angle between the front wheel of the bicycle and the longitudinal
axis of the car. Both the curvature and the curvature derivative are limited, i.e.
||
max
, ||
max
. (3)
To design the tracking controller we use a simpler version of (1) without the
curvature.
q =

x
y

cos
sin
tan
b

v. (4)
In case of the semi-autonomous system the controller cannot consider the
longitudinal velocity of the vehicle as a system input since it is generated by the
driver and cannot be inuenced by the controller. In this case we denote this
driver velocity by v
d
. (We suppose, that this v
d
velocity can be measured or well
estimated.)
q =

x
y

cos
sin
tan
b

v
d
. (5)
4 Path Planning
The task of the path planning method is to determine the geometry of the
reference path. Our goal is to have reference path with continuous curvature
which avoids stopping the car while steering the front wheels. Further constraints
are the maximal limit on the curvature and on its time derivative, since the
vehicle is not able to turn with arbitrary small turning radius and the change of
the turning radius (or the curvature) is also limited by the applied EPAS system.
To plan such a path which fullls the above mentioned constraints we use
three dierent path primitives namely straight lines, circular segments, and con-
tinuous curvature turns (CC turns) [11]. The curvature is zero along a straight
line, in a circular segment it has a nonzero constant value, which does not exceed
a given maximum limit, and the curvature varies linearly with the arc length in
the CC turns.
The geometry of the straight lines or circular segments can be described easily
if their parameters are known (e.g. lengths, turning radius). The calculation is
more complicated in the case of the CC turns. If the motion is started from the
initial conguration q
0
= [0, 0, 0, 0]
T
, the velocity prole is constant (e.g. v = 1),
and the curvature changes with the allowable maximum
max
value, then the
congurations in a CC turn can be described by the following equations:
x =

/
max
C
F

2
/(
max
)

, (6)
y =

/
max
S
F

2
/(
max
)

, (7)
=
2
/(2
max
), (8)
=
max
t, (9)
where C
F
and S
F
denote the Fresnel integrals, which cannot be given in a closed
form. Dierentiating (6-9) we can see that these equations fulll the kinematic
model of the vehicle given in (1).
Using simple mathematical operations for (6-9) we can get the congurations
for motions from dierent initial congurations with arbitrary constant velocity.
If the values of the Fresnel integrals in (6-7) can be calculated beforehand, then
the remaining computations can be performed in real time [16].
Now we describe how to use the three path primitives to get the reference
trajectory for parking in a lane. In this case we use seven segments to put the
path together (see Fig. 3). Without loss of generality we suppose that the vehicle
starts the motion backwards from the q
s
= [0, 0, 0,
s
]
T
initial conguration.
First it moves along a CC turn, until the maximal curvature (or the minimal
turning radius) is reached. Then it turns along a circle with the minimal turning
radius, the corresponding turning angle is denoted by
1
. After the next CC
turn the curvature becomes 0, and the car turns in the opposite direction with a
CC turn, a circular motion with turning angle
2
and one more CC turn, such
that the curvature becomes 0 again. The path ends with a straight line segment
whose length is denoted by l.
Such a path has three parameters: the turning angles in the circular segments
(
1
,
2
) and the length of the straight line (l). If these parameters are known,
the geometry of the path can be calculated and the goal conguration can be
determined:
q
g
=

x
g
y
g

f
1
(
1
,
2
, l)
f
2
(
1
,
2
, l)
f
3
(
1
,
2
)
0

. (10)
0
1 4
- 1
9
y
x R
CC
CC
CC
CC
circle
circle
straight
j
2
j
1
l
Fig. 3. Reference path for parking in a lane
The values of the path parameters can be determined from the desired
goal conguration (q
d
) by solving the q
d
= q
g
equation. So the full reference
path can be calculated, and the reference values for the conguration (q
ref
=
[x
ref
, y
ref
,
ref
,
ref
]
T
) and its time derivatives can be determined.
5 Time-Scaling
During the path planning we considered a preliminary reference velocity prole
and to avoid involved calculations we supposed that it is constant (e.g. v = 1).
Since the driver will generate another velocity prole (v
d
= v), it is enough if one
is able to track the geometry of the reference of the designed path and the time
distribution along the path will be adapted in real-time to the velocity prole
generated by the driver.
In the above equations the states of the conguration q were functions of
time t, where

t = 1. In a more general form we have a state equation
q(t) = f(q(t), u(t), w(t)), (11)
where u is the vector of the inputs and w denotes the external signals. In our
case, which is given by (5), we have u = and w = v
d
.
We introduce a new scaled time denoted by such that is used to modify
the time distribution along the path. We suggest that the relationship between
t and should not only depend on the states of the car (as it is usually made
in the literature [17]), but also on a new external input, denoted by u
s
which is
the so-called scaling input:
dt
d
=
1

= g(q, u
s
, w). (12)
Using this time-scaling (11) can be expressed with respect to the time :
q

=
dq
d
=
dq
dt
dt
d
= g(q, u
s
, w)f(q, u, w). (13)
The prime denotes dierentiation according to , hence

= 1.
The time-scaling dened in (12) has to satisfy some conditions:
(0) = t(0) = 0, since the original and the scaled trajectories should start
from the same initial conguration;
> 0, since time cannot stand or rewind.
During the time-scaling we modify the time distribution along the reference
path, which was planned in time :
x
ref
(t) = x
ref
(), (14)
x
ref
(t) = x

ref
() , (15)
x
ref
(t) = x

ref
()
2
+ x

ref
() . (16)
The further derivatives and the other state variables can be calculated in t in a
similar way. It can be seen from (14) that the time-scaling does not change the
geometry of the reference path, only the velocity and the further derivatives are
modied if = 1.
6 Tracking Control
In this section only some key features of the tracking controller are discussed,
the entire algorithm can be found in details in [14]. The literature suggests sev-
eral solutions [13, 18] to control the two-input kinematic car given in (4). These
methods ensure exponential tracking of the reference path. In our one-input case
these algorithms cannot be used without modication since our controller cannot
inuence the velocity of the vehicle. Our idea is to complement the lost velocity
input by the time-scaling input as in (13).
For, the following time-scaling function can be used:
dt
d
=
u
s
v
d
. (17)
In this case the model equation given in (5) which evolves according to t can be
transformed using the time-scaling, and we get
q

cos
sin
tan
b

u
s
. (18)
This scaled model has now two inputs (u
s
and ), hence one of the controllers
described in the literature can be used for tracking. The selected method will
compute and u
s
according to the tracking error between the real and the scaled
reference trajectories. This input is used to control the EPAS steering system
and the scaling input u
s
inuences the time-scaling. The time-scaling function
and its derivatives, which are required for (14-16), can be calculated using the
following relationships, which are based on (17)
(t) =

t
0
v
d
u
s
d, (0) = t(0) = 0, (19)
v
d
= u
s
, (20)
v
d
= u
s
+ u
s
. (21)
If the signs of u
s
and v
d
are the same than the time-scaling function satises the
> 0 condition. If one of the two signals equals 0, the car is not controllable.
This occurs at the very beginning and at the end of the motion.
The scheme of the closed loop control is depicted in Fig. 4. First the path
planning module calculates the reference path in . In the next step this refer-
ence is scaled based on the longitudinal velocity of the car v
d
, which is generated
by the driver, and based on the scaling input u
s
, which is calculated by the con-
troller. After the time-scaling we have the scaled reference in t. The controller
determines its outputs using the dierence between the real and the scaled ref-
erence trajectories. So the inputs of the vehicle are the longitudinal velocity v
d
generated by the driver, and , which is calculated by the controller. The output
of the car is the position of the reference point and the orientation.
( ) x y ,
y
d
car
driver
tracking
controller
time-
scaling
v
d
ref
in t
ref
in t
u
s
d
x y , ,y
motion
planning
Fig. 4. Scheme of the tracking controller with time-scaling
7 Results
We implemented the presented methods on a Ford Focus type passenger car (see
Fig. 5) using an EPAS provided by ThyssenKrupp to realize the steering angle
for the front wheels. Both the path planner and tracking controller were realized
Fig. 5. Ford Focus
on a dSPACE hardware (dedicated Autobox) mounted in the car and connected
to the EPAS actuator and to the CAN bus to read the ABS signals.
The presented motion is continuous curvature forward parking maneuver.
The results are depicted in Fig. 6. The system is able to track the reference
path, such that the velocity of the car is not constant, as it was supposed during
the path planning.
0 5 10 15
8
6
4
2
0
2
x [m]
y

[
m
]
Trajectories in the plane. Final position error: 0.021985 m


Reference initial position
Reference goal position
Real initial position
Real goal position
Reference
Real
0 2 4 6 8
0
1
2
3
time [s]
v
e
l
o
c
i
t
y

[
m
/
s
]
Inputs of the car
0 2 4 6 8
500
0
500
time [s]
s
t
w

a
n
g
l
e

[
d
e
g
]
Fig. 6. Results: path in the x y plane, inputs of the car v
d
,
8 Conclusions
This paper discussed two components of an APAS. The presented path planning
method can calculate a continuous curvature path in real time. The time-scaling
tracking controller is able to drive the car along the reference path such that the
driver generates the car velocity. The time-scaling function can be calculated
from the velocity of the car and from a scaling input, which is calculated by
the controller based on the closed loop behavior. The presented algorithms were
tested in a real car with encouraging results.
Acknowledgments. The research was partially funded by the Advanced Ve-
hicles and Vehicle Control Knowledge Center under grant RET 04/2004 and by
the Hungarian Science Research Fund under grant OTKA K71762.
References
1. Nissan Around View Monitor. http://www.nissan-
global.com/EN/TECHNOLOGY/INTRODUCTION/DETAILS/AVM/index.html
2. AISIN SEIKI Co. Ltd. Intelligent Parking Assist.
http://www.aisin.com/product/automotive/info/ot.html
3. Evolve. http://www.ikp.liu.se/evolve
4. Volkswagen Park Assist. http://www.volkswagen.com
5. Dubins, L.E.: On curves of minimal length with a constraint on average curvature,
and with prescribed initial and terminal positions and tangents. American Journal
of Mathematics 79(3) (1957) 497516
6. Reeds, J.A., Shepp, L.A.: Optimal paths for a car that goes both forwards and
backwards. Pacic Journal of Mathematics 145 (1990) 367393
7. Boissonnat, J.D., Cerezo, A., Lebond, J.: Shortest paths of bounded curvature in
the plane. Journal of Intelligent and Robotics Systems 11 (1994) 520
8. Laugier, C., Fraichard, T., Garnier, P.: Sensor-based control architecture for a
car-like vehicle. Autonomous Robot 6(2) (1999) 165185
9. Svestka, P., Overmars, M.H.: Motion planning for carlike robots using a proba-
bilistic learning approach. International Journal of Robotics Research 16(2) (1997)
119143
10. LaValle, S.M., Kuner, J.J.: Rapidly-exploring random trees: Progress and
prospects. In: Proceedings of the 2000 International Workshop on the Algorithmic
Foundations of Robotics. (2000) 293308
11. Fraichard, T., Scheuer, A.: From Reeds and Shepps to Continuous-Curvature
paths. IEEE Transaction on Robotics and Automation 20(6) (2004) 10251035
12. Zhao, Y., Collins, E.G., Dunlap, D.: Design of genetic fuzzy parallel parking control
systems. In: Proceedings of the American Control Conference. Volume 5. (2003)
41074112
13. Cuesta, F., Ollero, A.: Intelligent Mobile Robot Navigation. Volume 16 of Springer
Tracts in Advanced Robotics. Springer, Heidelberg (2005)
14. Kiss, B., Szadeczky-Kardoss, E.: Tracking control of the orbitally at kinematic
car with a new time-scaling input. In: Proceedings of the IEEE Conference on
Decision and Control. (to appear)
15. Paromtchik, I.E., Laugier, C.: Motion generation and control for parking an
autonomous vehicle. In: Proceedings of the IEEE International Conference on
Robotics and Automation. (April 1996) 31173122
16. Szadeczky-Kardoss, E., Kiss, B.: Continuous-curvature turns in motion planning
for car like mobile robots. In: Proceedings of the International Carpathian Control
Conference. (June 2007) 684687
17. Sampei, M., Furuta, K.: On time scaling for nonlinear systems: Application to
linearization. IEEE Transactions on Automatic Control 31 (1986) 459462
18. Fliess, M., Levine, J., Martin, P., Rouchon, P.: Flatness and Defect of Nonlinear
Systems: Introductory Theory and Examples. International Journal of Control
61(6) (1995) 13271361