Empirical Evaluation of Vehicular Models for Ego Motion Estimation
Robin Schubert, Christian Adam, Marcus Obst, Norman Mattern, Veit Leonhardt, and Gerd Wanielik
Abstract— Estimating the motion of a vehicle is a crucial
requirement for intelligent vehicles. In order to solve this
problem using a Bayes filter, an appropriate model of vehicular
motions is required. This paper systematically reviews typical
vehicular motion models and evaluates their suitability in dif-
ferent scenarios. For that, the results of extensive experiments
using accurate reference sensors are presented and discussed in
order to provide guidelines for the choice of an optimal model.
I. INTRODUCTION
One of the most promising technologies for increasing road
safety and driving comfort are advanced driver assistance
systems (ADASs). Such systems are characterized by the
perception of a vehicle’s surrounding using different, often
complementary sensors. In order to bridge the gap between
the inherent uncertainties of the sensor measurements and the
required reliability of the desired information, it is common
to apply statistical signal processing techniques.
One of the most widely used concepts is the Bayes filter
or, more precisely, one of its practical implementations. Each
Bayes filter, however, requires an appropriate probabilistic
state space model of the system under consideration which
consists of two parts: While the system or process model de-
scribes the temporal behavior, the sensor model characterizes
the properties of the utilized sensors [1].
For vehicular applications, a system model which accu-
rately represents the motion of a vehicle is often a crucial
requirement. While this is rather obvious for applications
which explicitly address the estimation of a vehicle’s motion
(e.g., systems for localizing the ego vehicle [
2
] or tracking
vehicles in the surrounding [
3
]), it is also true for non-
vehicular tracking applications. For instance, pedestrian
detection systems require an accurate ego motion estimation
in order to distinguish the motions of the pedestrians from
those of the ego vehicle [4].
Consequently, ego motion estimation is a part of virtually
any work related to the perception of a vehicle’s surrounding.
Though a number of different models have been proposed,
a systematic empirical evaluation is rarely available. In [
5
],
three of the most common models have been compared based
on simulated data. In [
6
], two models have been evaluated
for a mobile robot. Finally, an evaluation of sophisticated
models for navigation purposes has been presented in [7].
Another related work, which was done by some of the
authors, attempted to evaluate different motion models for
navigation purposes using highly accurate reference sensors
All authors are with the Professorship of Communications Engineer-
ing, Chemnitz University of Technology, Chemnitz, Germany, e-mail:
firstname.lastname@etit.tu-chemnitz.de
[
8
]. However, there are some factors which limit the signifi-
cance of this evaluation for ADAS applications. In contrast
to usual tracking algorithms, position measurements from a
global navigation satellite system (GNSS) have been used.
Furthermore, in lack of a more appropriate procedure, the
filter parameters have been tuned manually.
In this paper, the most common vehicular motion models
shall be empirically evaluated, in particular regarding their
suitability for ADAS applications. The basic idea is to
apply identical filters each of which contains a separate
motion model using automatically optimized parameters. The
experiments are conducted in different environments such as
highways or urban roads using a highly accurate reference
sensor system.
The paper is structured as follows: Section II introduces the
utilized notation and gives a brief overview about probabilistic
filtering and common motion models. In the subsequent
section III, the experimental setup as well as the evaluation
methodology are presented. Section IV contains detailed
quantitative results of the motion model evaluation. The
paper concludes with a discussion and recommendations
about appropriate models in section V.
II. FUNDAMENTALS
A. Probabilistic Filtering
The aim of Bayesian tracking algorithms is to recursively
estimate the probability density function (PDF) of a system’s
n
x
-dimensional state vector
x
k
∈ R
n
x
for each time step
k
.
The dynamic behavior of the system is represented by the
discrete-time stochastic model
x
k+1
= g
a
(x
k
, v
k
, T
k
), (1)
where
g
a
is called state transition equation. The vector
v
k
denotes an independent and identically distributed (i.i.d.)
system noise process, which represents the non-deterministic
part of the model.
T
k
denotes the time between
k
and
k + 1
.
In the general case, the state vector
x
k
is not observable;
however, it is assumed that measurements
y
k
∈ R
n
y
are
available which are connected to
x
k
by a known mathematical
relation, called the measurement equation
y
k
= g
c
(x
k
, w
k
), (2)
where w
k
denotes an i.i.d. observation noise time series.
For the filtering problem, the noise processes are assumed
to be white and mutually independent. Furthermore, the
sequence of all measurements which are available at time
k
is
denoted by
Y
k
b= {y
i
, i = 1, . . . , k}
. If the initial PDF
p(x
0
)
is assumed to be known, the filtering problem can be defined
as recursively determining the posterior PDF p(x
k
|Y
k
).
2011 IEEE Intelligent Vehicles Symposium (IV)
Baden-Baden, Germany, June 5-9, 2011
978-1-4577-0891-6/11/$26.00 ©2011 IEEE 534