Vehicle System Dynamics

ISSN: 0042-3114 (Print) 1744-5159 (Online) Journal homepage: http://www.tandfonline.com/loi/nvsd20

Dual extended Kalman filter for vehicle state and

parameter estimation

T. A. Wenzel , K. J. Burnham , M. V. Blundell & R. A. Williams

To cite this article: T. A. Wenzel , K. J. Burnham , M. V. Blundell & R. A. Williams (2006) Dual
extended Kalman filter for vehicle state and parameter estimation, Vehicle System Dynamics, 44:2,
153-171, DOI: 10.1080/00423110500385949

To link to this article: https://doi.org/10.1080/00423110500385949

Published online: 16 Feb 2007.

Submit your article to this journal

Article views: 1743

View related articles

Citing articles: 128 View citing articles

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=nvsd20
Vehicle System Dynamics
Vol. 44, No. 2, February 2006, 153–171

Dual extended Kalman filter for vehicle state and

parameter estimation
T. A. WENZEL*†, K. J. BURNHAM†, M. V. BLUNDELL† and R. A. WILLIAMS‡
†Control Theory and Applications Centre, Coventry University, Priory Street, CV1 5FB, Coventry, UK
‡Jaguar and Land Rover Research, Whitley, UK

The article demonstrates the implementation of a model-based vehicle estimator, which can be used
for combined estimation of vehicle states and parameters. The estimator is realised using the dual
extended Kalman filter (DEKF) technique, which makes use of two Kalman filters running in parallel,
thus ‘splitting’ the state and parameter estimation problems. Note that the two problems cannot be
entirely separated due to their inherent interdependencies. This technique provides several advantages,
such as the possibility to switch off the parameter estimator, once a sufficiently good set of estimates
has been obtained. The estimator is based on a four-wheel vehicle model with four degrees of freedom,
which accommodates the dominant modes only, and is designed to make use of several interchangeable
tyre models. The paper demonstrates the appropriateness of the DEKF. Results to date indicate that
this is an effective approach, which is considered to be of potential benefit to the automotive industry.

Keywords: Vehicle dynamics control; Dual extended Kalman filter; State and parameter estimation;
Modelling and simulation technology

1. Introduction

The automotive industry has made significant technological progress over the last decade or
so concerning active vehicle stability control, and hence improved safety, by developing and
introducing microprocessor control systems. Among these controllers are systems such as
the anti-lock braking system, active roll control, active front steering and electronic stability
programs [1]. Effective operation of each of these systems depends on an accurate knowledge
of the vehicle states, such as velocity, lateral acceleration, yaw rate, as well as vehicle and tyre
side slip. Due to physical and economic reasons, only a few of these states can actually be
obtained by measurement, thus the knowledge of the majority of these states is required to be
obtained by online estimation. As the control systems currently operate independently, their
functions are based on their individual vehicle state estimations. For a vehicle manufacturer,
it may be more desirable to develop a coordinated approach, overseeing these control systems
by designing a single state estimator, which is common to all systems, and hierarchically
configured to provide the necessary states.

*Corresponding author. Tel.: +44-2476-888972; Email: [email protected]/[email protected]

Vehicle System Dynamics

ISSN 0042-3114 print/ISSN 1744-5159 online © 2006 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/00423110500385949
154 T. A. Wenzel et al.

Accurate knowledge of vehicle states is essential and, not surprisingly, there are different
approaches for their estimation. However, most of these approaches are not without their
restrictions and limitations. Many of the estimators work only with a reduced number of states
or a reduced vehicle model, such as a bicycle model. Approaches can be found that use this
model for the estimation of side slip angle and yaw rate [2] or for lateral acceleration, yaw
rate and tyre slip angles [3, 4].
Other approaches make use of more comprehensive vehicle models, such as a four-wheel
model. For example, in ref. [5], use is made of a four-wheel model to estimate lateral velocity
and yaw rate, as well as the vehicle’s roll motion, but it has the drawback that the implemented
estimator requires the use of a whole set of additional costly acceleration sensors. A four-wheel
model is also used by Huh et al. [6] to estimate longitudinal and lateral velocity and yaw and
roll motion.
All the above approaches have another common limitation. They use fixed vehicle param-
eters, such as mass, moments of inertia and/or position of the centre of gravity (cog), which
are assumed to be known. While approximate values for these parameters are known, such as
the weight of an empty vehicle, these parameters may vary significantly from one journey to
the next. For example, when comparing a vehicle occupied only by the driver with a vehicle
with passengers, additional luggage and a full fuel tank, then the mass could easily vary by
several hundred kilograms. There are approaches to estimate some of the parameters, such as
mass [7], but these depend on the existence of further measured states. It can be shown that
appropriate knowledge of these vehicle parameters is of importance for effective estimation of
the various vehicle states. Recognising the need for such knowledge, the estimator described
in this paper additionally estimates the parameters to achieve an improved estimation of the
vehicle states.

2. Vehicle model

The state and parameter estimator presented here is based on a four-wheel vehicle model,
which comprises four degrees of freedom: the motion in the longitudinal direction x, i.e. the
longitudinal velocity vx ; the motion in the lateral direction y or lateral velocity vy ; yaw around
the vertical axis z, described by the yaw rate ψ̇ and roll with regard to the longitudinal axis x,
i.e. the roll angle φ.
Other states that depend directly on these four states can be directly derived, such as longi-
tudinal and lateral accelerations ax and ay , or depend indirectly on these states, such as body
slip angle β and torque  around the z-axis. The tyre states, such as wheel slip angle αij and
slip rate sij , as well as the wheel forward velocities vij and rotational velocities ωij are also
of importance. Note that for the states concerning the tyres, the indices ij represent the tyre
location on the vehicle, where i indicates the front (f) or rear (r) tyre and j indicates the left
(l) or right (r) side.
The vehicle states are largely dependent on the key operating parameters of a vehicle. Of
particular importance here are the vehicle mass m, moments of inertia Jz and Jsx and the
distances b and c between the centre of gravity and the front and rear axles, respectively.
Figure 1 illustrates the main states and parameters for a vehicle.
The interrelationships between the different vehicle dynamic states can be described by
differential equations. Various approaches for formulating these equations can be found in the
standard literature. The vehicle model implemented here utilises standard formulae, which
can be obtained from refs. [8, 9]. The differential equations for the calculation of longitudinal
 KF for vehicle state and parameter estimation 155

Figure 1. Illustration of vehicle states and parameters.

and lateral acceleration are as follows:

1
ax = (Fxfl cos δ − Fyfl sin δ + Fxfr cos δ
m
− Fyfr sin δ + Fxrl + Fxrr ) (1)
1
ay = (Fyfl cos δ + Fxfl sin δ + Fyfr cos δ
m
+ Fxfr sin δ + Fyrl + Fyrr ) (2)
v̇x = ax + vy ψ̇ (3)
v̇y = ay − vx ψ̇ (4)

Yaw and roll motion can be obtained from:


ψ̈ = (5)
Jz
tf
tf
tr tr
 = Fxfl − Fxfr + Fxrl − Fxrr
2 2 2 2



+ bFyfl + bFyfr − cFyrl − cFyrr
+ Mzfl + Mzfr + Mzrl + Mzrr (6)
1

φ̈ = −ms hs ay + φ(ms ghs − κφ ) − φ̇βφ (7)
Jsx

where the F
s denote the tyre forces acting in the vehicle coordinate system as follows:


Fxij = Fxij cos δ − Fyij sin δ (8)


Fyij = Fyij cos δ + Fxij sin δ (9)
156 T. A. Wenzel et al.

with tf and tr denoting the front and rear track widths, respectively. For the roll motion only
the sprung mass is taken into account, where ms denotes the sprung mass, Jsx the moment of
inertia around the longitudinal axis and hs denotes the height of the cog. The constants κφ and
βφ denote the roll stiffness and roll damping, respectively.
The longitudinal and lateral velocities, the steer angle of the front wheels and the yaw rate
are then used as a basis for the calculation of the tyre slip angles αij as well as the vehicle
body slip angle β:
 
vy + bψ̇
αfl,fr = δ − arctan (10)
vx ± tf ψ̇/2
 
−vy + cψ̇
αrl,rr = arctan (11)
vx ± tr ψ̇/2
 
vy
β = arctan (12)
vx
The slip rates of the tyres are calculated based on the ratio between the rotational wheel
velocities ωij and the actual velocities vij of the wheel centres on the road, following the
approach in ref. [10]:
ωij r
sij = −1 (13)
vij
where
 
tf
vfl,fr = vcog + ψ̇ ± − bβ (14)
2
 
tr
vrl,rr = vcog + ψ̇ ± + cβ (15)
2
in which vcog denotes the absolute velocity of the cog.
Due to longitudinal and lateral acceleration of the vehicle, the load distribution can vary
significantly during a journey. The load distribution can be expressed by the vertical forces
that act on each of the four wheels. These can be calculated as follows:
 
1 ay h c 1 h
Fzfl,zfr = mg ± m − max (16)
2 tf 2
 
1 ay h b 1 h
Fzrl,zrr = mg ± m + max (17)
2 tr 2
The lateral acceleration used in equations (16) and (17) is that which occurs at the cog. It
should be noted that, in general, this cannot be measured, because an accelerometer cannot
normally be placed in the exact cog, due to design constrictions. Also, the position of the
cog will vary depending on the actual load distribution of the vehicle. Consequently, the
measured and the actual acceleration will differ due to yaw and roll movement, depending on
the actual sensor position. A simplified approach for compensating for this discrepancy can
be found in ref. [11], which accommodates the additional acceleration measurement due to
both gravitational acceleration g and yaw acceleration ψ̈:

ay,sensor = (ay + ψ̈xa ) cos φ + g sin φ (18)

where xa is the longitudinal distance between the accelerometer and the cog. However, not
only the roll angle φ, but also the effects of roll rate φ̇ and roll acceleration φ̈ as well as
 KF for vehicle state and parameter estimation 157

yaw rate are required to be considered. Taking into account the accelerometer position with a
height hz above the roll axis, the longitudinal distance xa and lateral distance ya from the cog
and making use of the basic kinematic equations, which may be found in ref. [12], a refined
expression (superceding equation (18)) for the measured lateral acceleration can be deduced:


ay,sensor = ay + ψ̈xa − ψ̇ 2 ya + φ̈hz − φ̇ 2 ya · cos φ + g sin φ (19)

This can be simplified by using the approximations sin φ ≈ φ and cos φ ≈ 1 for small angles
of φ. Furthermore, assuming a lateral position of the accelerometer over the vehicle’s centre
line along the x-axis, and defining the longitudinal position of the accelerometer as ba behind
the front axle, equation (19) can be simplified to:


ay,sensor = ay + ψ̈(b − ba ) + φ̈hz + gφ (20)

3. Tyre model

The whole motion of the vehicle is a direct result of the forces (neglecting the aerodynamic
forces) that are generated between the road and the four tyres. The dynamic equations outlined
above show that all states depend on the calculation of the vehicle accelerations, which in turn
depend very much on accurate knowledge of the tyre forces, as can be seen in equations (1),
(2) and (6). Because the actual tyre forces cannot be measured, they need to be calculated
based on the estimated and measured states.
There are different approaches for achieving this, such as the so-called ‘Magic Formula’
by Pacejka [13], the tyre model by Fiala [14] or the ‘TMeasy’ tyre model [15]. The estimator
described in this article is configured such that use may be made of any suitable tyre model,
thus facilitating comparison between the various approaches.
Essentially, the tyre models differ in terms of their relative complexities, i.e. number of
coefficients, which are specific for each particular model. The number of coefficients can vary
widely, e.g. from 9 in the Fiala model to 82 in the latest version of the Magic Formula. These
parameters are normally obtained by extensive experimental testing of tyres on a test rig.
The tyre models require certain common inputs, such as tyre slip angle, tyre slip rates and
vertical forces acting on each tyre. The example presented in section 5 makes use of both
the Magic Formula and TMeasy for determining the tyre forces. The Magic Formula is used
to calculate the longitudinal and lateral tyre forces Fx and Fy as well as the self-aligning
moments Mz by a scheme, which is given by the following equations [16]:

y = D sin[C arctan{Bx − E(Bx − arctan Bx)}] (21)

Y (X) = y(x) + Sv (22)
x = X + Sh (23)

where the general output variable Y in equation (22) is used to represent the tyre forces Fx
and Fy and self-aligning moments Mz , while the input variable X is used to represent tyre
slip rate s (when calculating Fx ) or tyre slip angle α (when calculating Fy and Mz ). The main
issue is then the determination of the coefficients B, C, D, E, Sv and Sh , which are calculated
using the vertical tyre force Fz , the road friction µ and the specific tyre model parameters.
Note that y and x in equations (21)–(23) are merely intermediate variables in determining the
general output Y (X).
In contrast to this approach, which calculates each directional force separately, the TMeasy
tyre model calculates the resulting force F on the tyre by a combination of rational and
158 T. A. Wenzel et al.

polynomial functions, as can be found in ref. [15]. Using the same notation as in ref. [15], the
force on the tyre as a function of slip is given by:
 s

 s M dF 0 1+σ (σ +dF 0σ(s M /F M )−2) , σ = M , 0 ≤ s ≤ sM


 s
F (s) = F M − (F M − F G )σ 2 (3 − 2σ ), σ = s − s , s M < s ≤ s G
M
(24)

 G − sM

 s
F G , s > sG

where s denotes the resulting slip. The maximal possible resulting force is denoted by F M
and the resulting force in full sliding condition is denoted by F G . The corresponding slip rates
are denoted by s M and s G , respectively. The actual values for the forces F M and F G and the
slip rates s M and s G are obtained by performing quadratic interpolation, depending on the
vertical force, using values specified in look-up tables. Furthermore, dF 0 denotes the overall
initial inclination of the horizontal tyre forces. The resulting force F may be decomposed into
longitudinal and lateral forces:

Fx = F cos α (25)
Fy = F sin α (26)

Despite the complexity of the tyre models, there are several reasons why such models do
not match the actual tyre forces perfectly [17]. For example, the rig testing may have taken
place under well behaved conditions corresponding to operation in a linear region, while the
tyres may find themselves in rapidly changing conditions and may even experience effects
of non-linear behaviour at the limit of the linear region. In addition, although representative
surfaces may be used for testing, these surfaces can differ significantly from the road surface on
which the tyres are subsequently used. Thus, when choosing and using a tyre model, due care
needs to be given, recognising that whilst a model may be quite advanced, it will nevertheless
introduce an element of uncertainty into the overall vehicle model.

4. Dual extended Kalman filter

The state and parameter estimator proposed in this article is based on the Kalman filtering
technique, which is a widely used conceptual two-stage prediction/correction approach that
has found application in a variety of areas. Whilst the Kalman filter (KF) was originally
proposed by Kalman [18] for the estimation of the non-measurable states in linear systems,
it has found wide usage in its extended form for state estimation of non-linear systems as
well as joint state/parameter estimation for linear and non-linear systems. The fundamental
operation of the KF is a successive process of prediction based on system input, followed by
the correction based on measurable system output.
A powerful extension, developed for state estimation in non-linear systems, is the so-called
extended Kalman filter (EKF), in which the state equations are linearised at each working point
[19]. The main concept is the same as the standard KF, but the computational complexity is
enlarged, as the linearisation requires the evaluation of a Jacobian matrix at each time step.
Recognising that the resulting augmented system is in fact non-linear, such an approach has
also been developed for joint state/parameter estimation for linear systems.
A further development also designed for state and parameter estimation is the dual extended
Kalman filter (DEKF), which was proposed by Wan and Nelson [20]. This variant provides
a ‘boot-strapping’ procedure for combined state and parameter estimation using two EKFs in
 KF for vehicle state and parameter estimation 159

parallel. Figure 2 illustrates the working scheme of the DEKF, consisting of alternating steps
of prediction followed by correction.
The vehicle parameters, such as mass, can be assumed to be constant during a journey or
vary at least only marginally, e.g. small reduction due to consumption of fuel. This indicates
an advantage of using two EKFs instead of a single EKF, which combines both state and
parameter vectors. Using the DEKF, it is possible to switch off the parameter estimator, once
a sufficiently good set of estimates for the parameters have been found. This should increase
the performance of the state estimator, as it reduces the parameter uncertainties as well as
disturbances arising from the varying model parameters.
In general a non-linear system can be formulated as:

xs (t + 1) = f (xs (t), u(t), xp (t), w(t)) (27a)

y(t) = h(xs (t), v(t), xp (t)) (27b)

where xs is the state vector, xp is the parameter vector, u is the input vector, y is the output
vector, with w and v being the process noise and output noise vectors respectively. The basic
equations for the DEKF for such a non-linear system can be found in [20]; they are stated here
as follows:

Parameter prediction:

x̂p− (t) = x̂p (t − 1) (28)

−
p (t) =
−
p (t − 1) + Rp (29)

Figure 2. Scheme of DEKF.

160 T. A. Wenzel et al.

State prediction:

x̂s− (t) = f (x̂s− (t − 1), u(t), x̂p− (t)) (30)

−
s (t) = Js (t)
−
s (t − 1)JsT (t) + Rs (31)

State correction:

Ks (t) =
− − T −1
s (t)Hs [σs + Hs
s (t)Hs ]
T
(32)
x̂s (t) = x̂s− (t) + Ks (t)[y(t) − Hs x̂s− (t)] (33)

s (t) = [I − Ks (t)Hs ]
−
s (t) (34)

Parameter correction:

Kp (t) =
− − T −1
p (t)Hp [σp + Hp
p (t)Hp ]
T
(35)
x̂p (t) = x̂p− (t) + Kp (t)[y(t) − Hs x̂s− (t)] (36)

p (t) = [I − Kp (t)Hp ]
−
p (t) (37)

Here, Rs and Rp are user-specified process noise covariance matrices for the state and
parameter estimators, respectively; σs and σp are the corresponding output noise covariance
matrices, respectively, and
s and
p are the covariance matrices of the estimation errors,
respectively.
Using equations (1)–(20) for defining the system dynamic equations f (·) and h(·), the
Jacobian matrices Js and Hs for the state and output equations are then given, respectively, by:
 
∂f1 ∂f1
···
 ∂x1 ∂xm 
 . .. 
 . 
Js =  . .  (38)
 
 ∂fm ∂fm 
···
∂x1 ∂xm
 
0 0 1 0 ··· 0
Hs = 0 0 0 0 · · · 1 (39)
1 0 0 0 ··· 0

The input vector u and the output vector y consist of the available measurable vehicle states:

δ
 ωfl 
 
u= ωfr 
 (40)
ωrl 
ωrr
 
ψ̇
y = ay,sensor  = Hs xs (41)
vx
 KF for vehicle state and parameter estimation 161

The state and parameter vectors comprise the internal states required by the dynamic safety
control systems and the unknown vehicle parameters, respectively:

xs = [vx vy ψ̇ ay  β ax αfl αfr αrl αrr

Fz,fl Fz,fr Fz,rl Fz,rr sfl sfr srl srr φ φ̇ ay,sensor ]T (42)
 
m
 
xp = Jz  (43)
b

Note that the matrix Hp is required, which can be simplified as follows [21]:

∂f (x̂s , x̂p )
Hp = Hs (44)
∂ x̂p
 
∂ ψ̇ ∂ ψ̇ ∂ ψ̇
 
 ∂m ∂Jz ∂b 
 ∂ay,sensor ∂ay,sensor ∂ay,sensor 
=
 ∂m
 (45)
 ∂Jz ∂b  
 ∂vx ∂vx ∂vx 
∂m ∂Jz ∂b
Equations (28)–(37) depend on the definition of the four covariance matrices. Since for
both parallel estimators, the same output y is used, the measurement noise covariances are the
same: σs = σp , denoted σ. Thus the number of covariance matrices is reduced to three.
According to ref. [22], the measurement noise covariance matrix σ can be obtained via
off-line sample measurements, while it is more difficult to define the process noise covariance
matrix. In the case of state estimation, the estimator works well with a fairly ‘crude’ initial
setting, such as setting the process noise covariance matrix to the identity matrix with the same
constant value R0 for the diagonal elements. Experiments show that setting R0 to a sufficiently
large value, such as R0 = 10,000, results in good state estimates. In this example, good results
for the state estimation were obtained with the following settings:
 2 
sψ̇ 0 0
 
 2 
σs = σp =  0 say,sensor 0  (46)
 2

0 0 svx

Rs = R0 I (47)

However, the parameter estimation is found to be more sensitive to the actual process noise
covariance matrix settings. The initial covariance matrix for the parameter estimator was set
to the following diagonal matrix:
 2 
sm 0 0
 
Rp =  2
 0 s Jz 0 
 (48)
2
0 0 sb

For the following example the process noise covariance matrix Rp was set such that the
variables sm , sJz and sb were of an order of approximately 1% of the corresponding actual
values for the parameters m, Jz and b.
162 T. A. Wenzel et al.

The better the parameter estimation becomes, the smaller become the model uncertainties.
This can be taken into account by weighting the covariance matrices exponentially. Thus, in a
similar manner as in ref. [23], the covariance matrices corresponding to the model uncertainties
are decreased at each time step by a factor γ < 1:

Rs (t) = γ Rs (t − 1) (49)
Rp (t) = γ Rp (t − 1) (50)

A further feature of the DEKF implemented here is the possibility of complete and partial
switch off. For example, during vehicle stand-still, noise would be the only input to the DEKF;
therefore, both state and parameter estimation schemes are switched off. Consequently, the
DEKF is deactivated whenever zero velocity is approached.
Additionally, when the output signals corresponding to yaw rate, lateral acceleration and
longitudinal velocity are below certain thresholds, implying that the signal information is less
than the noise, then the parameter estimator is also switched off and the previous parameter
estimates are maintained.

5. Results

The DEKF algorithm used here is implemented in the industry standard MATLAB® [24]
functional programming environment.
Available data for a rearwheel drive executive saloon car was used to demonstrate the efficacy
of the estimator. The data sets comprise the vehicle parameters and the dynamic vehicle states
obtained both from an independent simulation using an alternative software package as well
as actual measured data from a single test drive using an actual vehicle.
The alternative simulated data was generated by the simulation software-package
ve-DYNA® [25]. The advantage of using simulated data, e.g. ve-DYNA, as a comparison is
the possibility to obtain the non-measurable states for reference, as shown initially in ref. [26].
Figure 3 shows that the DEKF works well for state and parameter estimation for such simulated,
noise-free data. Further results contained in this article confirm these preliminary findings and
also show that the DEKF estimator is capable of handling realistic data from an actual test
drive in the presence of measurement noise.
Data from the single test drive provides a realistic data set, e.g. with combined braking and
cornering, as well as measurement noise, similar to that which can be expected in a practical
application.
The example uses data that were obtained by driving on a public road at low speed. The
manoeuvres carried out consisted of acceleration, braking and cornering. Data were measured
over a time span of 135 s with a sampling period of 50 ms. Figure 4 shows the steer angle δ
and the wheel velocities corresponding to the rotational velocities ωij , which were used as
inputs for the DEKF. Figure 5 shows the reference outputs forward velocity vx , yaw rate ψ̇
and measured lateral acceleration ay,sensor .
Data for the tyres used on the vehicle – the parameters for the Magic Formula tyre model –
were available. Parameters for slightly different tyres for the TMeasy tyre model were also
available. This deliberate discrepancy provided the possibility to investigate the sensitivity of
the DEKF to the tyre model, including the accuracy of such parameters. The estimator was
then run with both tyre models. Figure 6 shows the estimated parameters: mass m, moment of
inertia Jz and position b of the cog when using the tyre models, namely, Magic Formula and
TMeasy. It can be seen that the Magic Formula tyre model provides final estimates within a
range of about 4% of the actual values. However, with the same settings the use of the TMeasy
 KF for vehicle state and parameter estimation 163

Figure 3. Estimated states for noise-free simulation data.

tyre model provides worse estimates, giving a final error in terms of the estimated mass of the
order 20%.
The results for the state estimation using both the Magic Formula tyre model and the TMeasy
tyre model are shown in figure 7, which depicts the estimates of the lateral velocity vy , the
lateral acceleration ay , vehicle slip angle β and the front left wheel tyre slip angle αfl , the
vertical force Fzfl and the slip rate sfl . The input data were also used within an independent

Figure 4. Input to DEKF.

164 T. A. Wenzel et al.

Figure 5. Reference outputs to DEKF.

Figure 6. Estimated parameters mass m, moment of inertia Jz and longitudinal position b of the cog.
KF for vehicle state and parameter estimation 165

Figure 7. Estimated states.

166 T. A. Wenzel et al.

Table 1. RMS error of selected states.

State Using TMeasy Using Magic Formula

vy (m s−1 ) 0.027668 0.032576

ay (m s−2 ) 0.23524 0.39377
β (rad) 0.013016 0.013524
αfl (rad) 0.0070118 0.0081403
Fzfl (N) 547.77 407.18
sfl (1) 0.014151 0.014447

reference vehicle simulation model to obtain reference values for the non-measurable states.
Comparing the state estimates from the DEKF, using both tyre models with the reference
simulation, it can be seen that the estimation is successful. By considering the root mean
square error (RMS) of the estimates over a period of 135 s versus the reference simulation as
given (see results given in table 1), it is found that the errors are acceptably small and appear
to be independent of the tyre model used. Only the estimation of the vertical force differs
appreciably from the reference data. This is believed to be caused by the strong dependence
of the vertical forces on the vehicle mass, which, for the time range shown, has not converged
sufficiently towards the actual values, but is known to differ by up to 8%. There appears to be
interdependence between the error in the mass estimation and the error in the vertical force
estimation. It can be seen that the DEKF is able to compensate for inaccurate vehicle and tyre
parameters in the case of the state estimator, thus demonstrating that it is able to provide fairly
good estimates for the main states.

Figure 8. Estimated parameters mass m, moment of inertia Jz and longitudinal position b of the cog for various
settings of the process noise covariance matrix Rp .
 KF for vehicle state and parameter estimation 167

Figure 9. Estimated states for various settings of the process noise covariance matrix Rp .
168 T. A. Wenzel et al.

Table 2. Root squared error of selected states for various

process noise covariance matrix Rp .

State DEKF1 DEKF2 DEKF3

vy (m s−1 ) 0.032576 0.032012 0.031152

ay (m s−2 ) 0.39377 0.37471 0.3481
β (rad) 0.013524 0.013451 0.01334
αfl (rad) 0.0081403 0.0080705 0.0079272
Fzfl (N) 407.18 400.18 396.68
sfl (1) 0.014447 0.014413 0.014356

In contrast, it is found that the parameter estimator is significantly more sensitive than the
state estimator, both in terms of dependency on the accuracy of the vehicle and tyre models
used and of the noise. From the results obtained, the importance of using the correct tyre
data is clearly evident. Inaccurate coefficients for the tyre model result in large errors for the
parameter estimation.
It is instructive to note the influence of the process noise covariance matrices on the DEKF.
The results obtained when running the estimator three times, using the Magic Formula only,
with initial process noise covariance matrix Rp given by equation (48) as well as 2Rp or 3Rp ,
are stored in the data sets DEKF1, DEKF2 and DEKF3, respectively. Figure 8 shows the
influence of the choice of Rp on the parameter estimation. However, the final estimates are
still within a range of approximately ±6% for all three estimations.
Figure 9 shows that the state estimation is very similar for all three cases. The RMS error,
given in table 2, shows again that the DEKF is capable of compensating for incorrect parame-
ters. It can be concluded that the selection of the process noise covariance matrix is a significant
factor for the parameter estimation.

6. Conclusions

This article has demonstrated the application of the DEKF technique for realising a model-
based vehicle state and parameter estimator. This technique makes use of two parallel driven
EKFs, which interact in a ‘boot-strapping’ way. Whilst one of the two EKFs is configured
for the estimation of the main vehicle parameters, the other EKF is configured for estimating
the vehicle states. One advantage of this technique is that with an efficient estimator and an
improved accuracy of the parameter estimates, the model uncertainties within the state esti-
mator decrease, thus allowing an enhancement of the quality of the state estimation. A further
benefit is the possibility to switch off the parameter estimator, once the parameter estimates
converge to satisfactory values or when the signal/noise ratio indicates disturbed input and
output signals. In this case, constant parameter estimates are maintained, thus reducing the
uncertainties for the state estimator.
The estimator is implemented in the Matlab/Simulink software environment and is based
on a four-wheel dynamic vehicle model with different interchangeable tyre models. The latter
feature offers the added flexibility to investigate the sensitivity of the estimator when using
different tyre models.
Evaluation of the state and parameter estimator with both independently simulated data
and measured data from a test drive using different tyre models shows that the state estimator
is reasonably robust in terms of coping with variations in the tyre model. It is able to com-
pensate for inaccuracies in the tyre parameters. However, the parameter estimator is much
more sensitive both in terms of the tyre models and in terms of the initial settings of the KF
 KF for vehicle state and parameter estimation 169

process noise covariance matrices. In general, the results to date are promising and show the
effectiveness of this approach. The principle of replacing the existing individual state estima-
tors, required by the various stability control systems within a vehicle, by one comprehensive
single state and parameter estimator has been demonstrated, and potential benefits have been
highlighted.
Further work is planned to extend the estimator by increasing the number of states and
parameters, whilst at the same time improving stability and computational efficiency. Of
particular interest is the inclusion of the vehicle’s environment, such as the road friction µ,
which could be treated as an additional state to be estimated.

References
[1] Bauer, H. (Ed.), 2000, Automotive Handbook (5th edn) (Robert Bosch GmbH: Stuttgart).
[2] Bolzern, P., Cheli, F., Falciola, G. and Resta, F., 1999, Estimation of the non-linear suspension
tyre cornering forces from experimental road test data. Vehicle System Dynamics, 31(1), 23–34.
doi:10.1076/vesd.31.1.23.2100.
[3] Venhovens, P.J.T. and Naab, K., 1999, Vehicle dynamics estimation using Kalman filters. Vehicle System
Dynamics, 32(2–3), 171–184. doi:10.1076/vesd.32.2.171.2088.
[4] Zuurbier, J. and Bremmer, P., 2002, State estimation for integrated vehicle dynamics control. Proceedings of
the 6th International Symposium on Advanced Vehicle Control (AVEC), Hiroshima, Japan.
[5] Satria, M. and Best, M.C., 2002, Comparison between Kalman filter and robust filter for vehicle handling
dynamics state estimation, SAE 2002 World Congress, Detroit, Michigan, 4–7 March, Warrendale: SAE Society
of Automotive Engineers. SAE technical paper series 2002-01-1185.
[6] Huh, K., Kim, J. and Yi, K., 2003, Monitoring system design for estimating the lateral tyre force. Proceed-
ings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering, 217, 247–256.
doi:10.1243/09544070360613219.
[7] Lingman, P. and Schmidtbauer, B., 2001, Road slope and vehicle mass estimation using Kalman filtering.
Proceedings of the 17th IAVSD Symposium in The Dynamics of Vehicles on Roads and on Tracks, Lyngby,
Denmark.
[8] Will, A.B. and Żak, S.H., 1997, Modelling and control of an automated vehicle. Vehicle System Dynamics,
27, 131–155.
[9] Wong, J.Y., 2001, Theory of Ground Vehicles (3rd edn) (John Wiley & Sons, Inc.: New York).
[10] Kiencke, U. and Nielsen, L., 2000, Automotive Control Systems (Springer: Berlin).
[11] Chee, W., 2001, Measuring yaw rate with accelerometers. Future Transportation Technology Conference, Costa
Mesa, California, SAE technical paper series 2001-01-2535.
[12] Stöcker, H. (Ed.), 1998, Taschenbuch der Physik (3rd edn) (Verlag Harri Deutsch: Thun und Frankfurt am Main).
[13] Pacejka, H.B. and Bakker, E., 1991, The magic formula tyre model. Proceedings of the 1st International
Colloquium on Tyre Models for Vehicle Dynamic Analysis (Swets & Zeitlinger: Amsterdam), pp. 1–18.
[14] Fiala, E., 1954, Seitenkräfte am rollenden Luftreifen. VDI Zeitschrift, 96(29), 973–979.
[15] Hirschberg, W., Rill, G. and Weinfurter, H., 2002, User-appropriate tyre-modelling for vehicle dynamics in
standard and limit situations. Vehicle System Dynamics, 38(2), 103–125. doi:10.1076/vesd.38.2.103.5620.
[16] Pacejka, H.B., 2002, Tyre and Vehicle Dynamics (Butterworth–Heinemann: Oxford).
[17] Lidner, L., 1991, Experience with the magic formula tyre model. Proceedings of the 1st International Colloquium
on Tyre Models for Vehicle Dynamic Analysis (Swets & Zeitlinger: Amsterdam), pp. 30–46.
[18] Kalman, R.E., 1960, A new approach to linear filtering and prediction problems. Transactions of the ASME–
Journal of Basic Engineering, Series D. 82, 35–45.
[19] Young, P.C., 1974, Recursive approaches to time series analysis. Bulletin of the Institute of Mathematics and its
Applications, 10, 209–224.
[20] Wan, E.A. and Nelson, A.T., 2001, Dual extended Kalman filter methods. In: S. Haykin (Ed.) Kalman Filtering
and Neural Networks, Chapter 5 (John Wiley & Sons: New York).
[21] Wan, E. and Nelson, A., 1997, Neural dual extended Kalman filtering: applications in speech enhancement and
monaural blind signal separation. Proceedings of Neural Networks for Signal Processing Workshop, IEEE.
[22] Welch, G. and Bishop, G., 2002, An Introduction to the Kalman Filter (University of North Carolina: Chapel
Hill).
[23] Sasiadek, J.Z. and Wang, Q., 2003, Low cost automation using INS/GPS data fusion for accurate positioning.
Robotica, 21, 255–260. doi:10.1017/S0263574702004757.
[24] The MathWorks, Inc., 2003, MATLAB. Available online at: http://www.mathworks.com. (accessed April 2003).
[25] TESIS, 2003, ve-DYNA, Gesellschaft für Technische Simulation und Software mbH. Available online at:
http://www.tesis.de (accessed October 2003).
[26] Wenzel, T.A., Burnham, K.J., Blundell, M.V. and Williams, R.A., 2004, Approach to vehicle state and parameter
estimation using extended Kalman filtering. Proceedings of the 7th International Symposium on Advanced
Vehicle Control AVEC04, HAN University, Arnhem, NL, 23–27 August, pp. 725–730.
170 T. A. Wenzel et al.

Nomenclature

ax longitudinal acceleration (m s−2 )

ay lateral acceleration (m s−2 )
ay,sensor measured lateral acceleration (m s−2 )
b distance between cog and front axle (m)
ba distance between accelerometer and front axle (m)
c distance between cog and rear axle (m)
f state equations
Fxij longitudinal forces on each wheel (N)
Fyij lateral forces on each wheel (N)
Fzij vertical forces on each wheel (N)
g gravitational acceleration (m s−2 )
h height of cog (m)
hs height of cog of sprung mass (m)
hz height of accelerometer above roll axis (m)
h output equations
Hs , Hp Jacobian matrix of output for state/parameter estimator
i front(f) or rear(r)
j left(l) or right(r)
Jsx moment of inertia of sprung mass around longitudinal axis (kg m2 )
Jz moment of inertia around vertical axis (kg m2 )
Js Jacobian matrices for state estimates
Ks , Kp Kalman gain matrices for state/parameter estimator
length of wheel base ( = b + c) (m)
m mass of vehicle (kg)
ms sprung mass (kg)
Mzij self-aligning moment off each tyre (N m)
r tyre radius (m)
Rs , Rp process noise covariance matrices for state/parameter estimator
sij wheel slip rate
tf , tr front and rear track width (m)
u input vector
vcog velocity of the centre of gravity (m s−1 )
vij forward wheel speed (m s−1 )
vx longitudinal velocity (m s−1 )
vy lateral velocity (m s−1 )
v output noise vector
w process noise vector
xs , xp state/parameter vector
y output vector
αij wheel slip angles (rad)
β body side slip angle (rad)
βφ roll damping constant (Nm s)
γ reduction factor for process noise covariance matrices
 KF for vehicle state and parameter estimation 171

δ steer angle (rad)

 torque around z-axis (N m)
κφ roll stiffness constant (N m)
σs , σp measurement noise covariance matrices for state/parameter estimator
φ body roll angle (rad)

s ,
p estimate error covariance matrices for state/parameter estimator
ψ̇ yaw rate (rad s−1 )
ωij wheel rotational velocity (rad s−1 )