Estimation of vehicle sideslip, tire force and wheel

cornering stiffness
Guillaume Baffet, Ali Charara, Daniel Lechner

To cite this version:

Guillaume Baffet, Ali Charara, Daniel Lechner. Estimation of vehicle sideslip, tire force and wheel
cornering stiffness. Control Engineering Practice, Elsevier, 2009, 17 (vol17, n11), pp1255-1264.
�10.1016/j.conengprac.2009.05.005�. �hal-00849264�

HAL Id: hal-00849264

https://hal.archives-ouvertes.fr/hal-00849264
Submitted on 30 Jul 2013

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
Control Engineering Practice, Vol 17, Issue 11, November 2009, Pp 1255-1264
doi:10.1016/j.conengprac.2009.05.005
Copyright © 2008 Elsevier Ltd All rights reserved.
http://www.sciencedirect.com/science/journal/09670661

Estimation of vehicle sideslip, tire force and

wheel cornering stiffness

Corresponding author G. Baffet, Tel. +33 344234645. Fax +33 344234477.

Guillaume Baffet a, Ali Charara a, Daniel Lechner b

a HEUDIASYC
Laboratory, UMR CNRS 6599,
Université de Technologie de Compiègne, Centre de recherche Royallieu,
BP20529 - 60205 Compiègne, France
b INRETS-MA Laboratory (Department of Accident Mechanism Analysis),
Chemin de la Croix Blanche, 13300 Salon de Provence, France

Abstract

This paper presents a process for the estimation of tire-road forces, vehicle sideslip
angle and wheel cornering stiffness. The method uses measurements (yaw rate, longi-
tudinal/lateral accelerations, steering angle and angular wheel velocities) only from
sensors which can be integrated or have already been integrated in modern cars.
The estimation process is based on two blocks in series: the first block contains
a sliding-mode observer whose principal role is to calculate tire-road forces, while
in the second block an extended Kalman filter estimates sideslip angle and cor-
nering stiffness. More specifically, this study proposes an adaptive tire-force model
that takes variations in road friction into account. The paper also presents a study
of convergence for the sliding-mode observer. The estimation process was applied
and compared to real experimental data, in particular wheel force measurements.
Experimental results show the accuracy and potential of the estimation process.

Keywords: Vehicle dynamics, Tire-road force, Sideslip angle, Cornering stiffness,

Sliding-mode observer

Email addresses: gbaffet@hds.utc.fr (Guillaume Baffet),

acharara@hds.utc.fr (Ali Charara), daniel.lechner@inrets.fr (Daniel
Lechner).

Preprint submitted to Elsevier 10 February 2009

1 Introduction

Knowledge of vehicle variables such as tire forces and sideslip angle is essential
for improving car safety, handling characteristics and comfort. Active safety
systems such as anti-lock braking systems and electronic stability programs
can significantly reduce the number of road accidents, and these safety sys-
tems may be improved if the variables of a car are well known. For instance,
information about road friction means a better definition of potential trajec-
tories, and therefore a better management of vehicle controls. However for
both technical and economical reasons some fundamental data (such as tire
forces, sideslip angle, road friction) are not measurable in a standard car. As
a consequence, tire forces and sideslip angle must be observed or estimated.
Fig. 1 illustrates wheel dynamic variables. The wheel sideslip angle represents
the angular direction of the linear wheel velocity relative to the longitudinal
wheel axis.

Vertical tire-force

Angular wheel
speed
Lateral
tire-force Road friction
Linear wheel
velocity
Wheel sideslip Longitudinal tire-force
angle

Fig. 1. Wheel dynamic variables.

Vehicle-dynamic estimation has been widely discussed in the literature, e.g.

(Arndt, 2006; Stephant, Charara & Meizel, 2007; Hsu & Gerdes, 2006; Gustafs-
son, 1997; Kim, 2009; Lakehal-ayat, Tseng, Mao & Karidas, 2006; Ono et al.,
2003; Tanelli, Savaresi & Cantoni, 2006; Wesemeier, 2008). The modeling of
vehicle-road systems is complex because the vehicle trajectory depends on
a wide variety of parameters, including tire, car and road properties. In the
literature, the vehicle-road system is usually modeled by combining a vehi-
cle model with a tire-force model (Baffet, Charara & Lechner, 2007; Lechner,
2002; Ungoren, Peng & Tseng, 2004). In this study, the estimation process is
separated into two blocks (Fig. 2), where block 1 corresponds to the car body
dynamic and block 2 corresponds to the tire-road interface dynamic.
The first block contains a sliding mode observer (denoted as SMO) which esti-
mates tire forces and yaw rate. This observer is constructed with a single-track

2
model (Segel, 1956), in which forces are modeled without tire-road parameters
(F˙x = 0, F˙y = 0). One advantage of this formulation is that the estimates will
not be influenced by tire-road parameters.
In the second block an Extended Kalman Filter (denoted as EKF) supplies
estimations of sideslip angle and wheel cornering stiffnesses. This observer is
developed from a sideslip angle model and a linear adaptive tire-force model.
The linear adaptive force model is proposed with an eye to correcting errors
resulting from road friction changes.
By separating the estimation process in two blocks, the force estimates, in
the first block, are not functions of road friction parameters. Consequently
the first observer is particularly robust with respect to road friction changes.
Moreover, the process is separated in two blocks for practical and analysis
reasons. In fact, the gains of the two observers are easier to adjust, than the
gains of one bigger observer combining the two observers.

Measurements: yaw rate, steering angle, lateral acceleration

longitudinal acceleration, angular wheel velocities
Block 1
Observer SMO
Sliding-mode observer
Single-track. model
.
Force model: Fx=0,Fy=0

Estimated variables: Longitudinal and lateral tire forces, yaw rate

Block 2
Observer EKF
Extended Kalman Filter
Sideslip angle model
Linear adaptive force model

Estimated variables: Sideslip angle and cornering stiffnesses

Fig. 2. The estimation process is divided into two blocks, where block 1 estimates
the tire-road forces and block 2 calculates sideslip angle.

The rest of the paper is organized as follows. Section 2 describes the modeling
of the first block, presents the sliding mode observer SMO and shows experi-
mental results for the estimates of tire forces. Section 3 describes the modeling
of tire-road forces, proposes the Extended Kalman Filter EKF, and presents
experimental results performed within the second block. The two observers are
evaluated with respect to sideslip angle and tire-force measurements. Section
4 describes the sensitivity to vehicle mass of the method. Section 5 presents
the conclusion. Appendix A analyzes the convergence properties of the sliding
mode observer SMO and Appendix B lists the different notations.

3
2 Tire-road force observer, SMO

2.1 Modeling of the sliding mode observer SMO

The observer SMO is based on the single-track model (Segel, 1956), which is
currently used to describe vehicle-dynamic behavior (Kiencke & Daib, 1997).
Fig. 3 presents the single-track model, where ψ̇ is the yaw rate, γx and γy are
the longitudinal and lateral accelerations, Vg is the vehicle velocity, β is the
sideslip angle, δ is the steering angle, Fyw1 and Fyw2 are the front and rear
lateral tire forces, Fxw1 and Fxw2 are the front and rear longitudinal tire forces,
and Fy1 and Fx1 are the front lateral and longitudinal tire forces in the car
body axis. L1 and L2 are the distances from the vehicle center of gravity to
front and rear wheels, x and y are the longitudinal and lateral vehicle posi-
tions, Vg1 and Vg2 the front and the rear vehicle velocities, and β1 and β2 the
front and rear sideslip angles.
Because of it simplicity, one advantage of the bicycle model is that time ex-

Fxw1 δ
Fyw1
δ
β1
y Vg1
x Fy1
Vg Fx1
γy β
γx
Fyw2 Fxw2 ψ
β2 L1
Vg2
L2

Fig. 3. The single-track model, also called the bicycle model, is used in the sliding
mode observer.

ecution is sufficiently short for real time applications. However, this model
ignores vehicle roll which has a large impact on the lateral acceleration, espe-
cially at high lateral accelerations.
In order to develop an observable system (notably in the case of null steering
angle) (Nijmeijer & Van der Schaft, 1990), rear longitudinal tire force Fxw2 is
neglected relative to the front longitudinal force Fx1 (assuming a front-wheel
drive car in braking situation). The simplified equation for yaw acceleration
can be formulated as the following dynamic relationship (single-track model):

1 (1)
ψ̈ = Iz
[L1 Fy1 − L2 Fyw2 ],

4
where Iz is the yaw moment of inertia. The force evolutions and the measure-
ments are modeled as:

Ḟy1 = 0, Ḟyw2 = 0, Ḟx1 = 0,

Fy1 + Fy2 Fx1 (2)
γy = , γx = ,
m m

where m is the vehicle mass. Consider the following state X∈R4 and mea-
surement Y∈R3 :

X = [x1 , x2 , x3 , x4 ] = [ψ̇, Fy1 , Fyw2 , Fx1 ],

(3)
Y = [y1 , y2, y3 ] = [ψ̇, γy , γx ].

Vectors x
b = [x
b1 , x
b2 , x
b3 , x c = [yb , yb , yb ] represent the state estimates
b4 ] and Y 1 2 3
and the measurement estimates. The measurement model is:
xb2 + xb3 xb4
yb1 = xb1 , yb2 = , yb3 = . (4)
m m

The estimation errors for states and measurements are denoted respectively
as:
f = [X − x
X b ] = [x
e1 , x
e2 , x
e3 , x
e4 ], f = [Y − y
Y b ] = [ye1 , ye2 , ye3 ]. (5)

The observer SMO uses a sliding mode structure based principally on the
works of (Drakunov & Utkin, 1995) and (Slotine, Hedrick & Misawa, 1987).
The state estimates evolve according to the single-track model (1), the force
model (2) and the sign of the measurement estimation errors:

1
xḃ1 = Iz
[L1 xb2 − L2 xb3 ] + Ω1 sign(ye1) + Ω2 sign(ye2 ) + Ω3 sign(ye3 ),
xḃ2 = Ω4 sign(ye1) + Ω5 sign(ye2 ) + Ω6 sign(ye3 ),
(6)
xḃ3 = Ω7 sign(ye1) + Ω8 sign(ye2 ) + Ω9 sign(ye3 ),
xḃ4 = Ω10 sign(ye1 ) + Ω11 sign(ye2 ) + Ω12 sign(ye3 ),

where Ω1 , ..., Ω12 are the observer gains. The convergence analysis of SMO is
described in appendix A. Choosing the observer gains such that:

{Ω2 , Ω3 , Ω6 , Ω9 , Ω10 , Ω11 } = 0,

{Ω1 , Ω4 , Ω5 , Ω8 , Ω12 } > 0, (7)
L2
Ω7 < 0, Ω4 = −Ω7 , Ω5 = Ω,
L1 8

5
implies the convergence of the estimation vector xb toward the system state X.
In the experimental tests the gains were set as:

• Ω1 = 10, Ω4 = 40000, Ω5 = 52000, Ω7 = −40000, Ω8 = 40000, Ω12 = 50000.

These gain values are chosen empirically, according to (7), numerous exper-
imental tests and in order that the estimates converge quickly to the states.
The sampling frequency is 100 Hz, and the chattering is avoided by using a
linear function (Levant (1993)).

Dynamometric hub Correvit sensor

Wheel-force transducers

Fig. 4. INRETS MA Laboratory’s experimental vehicle. Dynamometric hubs: wheel–

force transducers. Correvit sensors: sideslip angle and velocity transducers.

2.2 Experimental results - observer SMO

The INRETS MA vehicle (see Fig. 4) is a Peugeot 307 equipped with a number
of sensors including centimetric GPS, accelerometers, odometers, gyrometers,
steering angle, three Correvits and four dynamometric hubs. The three Corre-
vits (non-contact optical sensor) gives measurements of the sideslip angles and
velocities of the vehicle rear, the front-right wheel and the rear-right wheel.
The dynamometric hubs are placed at the four wheels and provide measure-
ments of tire forces and wheel torques.
This study included two experimental tests: a slalom and a roundabout. These
tests are representative of both longitudinal and lateral dynamic behaviors.
The vehicle trajectories and the acceleration diagrams are shown in Figs. 5
and 6.
These experimental tests were performed in dry asphalt. During the first test
(slalom), the vehicle first accelerated up to γx = 0.3 g (situation 1), then
negotiated a slalom (−0.6 g < γy < 0.6 g, situation 2), before finally deceler-
ating to γx ≈ −0.5 g (situation 3). In the second test the vehicle negotiated a

6
 First test: slalom

Lateral acceleration (m.s-2)

Lateral position (m)
1.5 5
1
2 1
0.5 0
0 3
-0.5
-1 -5
0 50 100 150 200 -4 -2 0 2
Longitudinal position (m) Longitudinal acceleration (m.s-2 )

Fig. 5. Slalom experimental test, vehicle position, acceleration diagram.

Second test: roundabout
Lateral acceleration (m.s-2)
Lateral position (m)

10
5
0
-10
-20 0
-30
-40
-5
0 20 40 60 -4 -2 0 2
Longitudinal position (m) Longitudinal acceleration (m.s-2 )

Fig. 6. Roundabout experimental test, vehicle position, acceleration diagram.

roundabout at maximum steering angle of 11o . The two diagrams show that
the vehicle is subject to lateral accelerations lower than 0.7 g.

The estimation results are presented in two formats: as numerical estimation

results compared with measurements, and as tables of normalized errors. The
normalized error εz for an estimation z is defined as:
100 · |z(i) − zmeas (i)|
εz (i) = . (8)
max |zmeas (i)|

This quantity is calculated for each sample. The vehicle parameters are Iz =
3000 kg.m−2 , L1 = 1.12 m, L2 = 1.46 m, m = 1447 kg. The estimates of
the tire forces along the wheel axis are obtained by applying the following
equations:

Fxw1 = Fx1 cos(δ) + Fy1 sin(δ), Fyw1 = Fy1 cos(δ) − Fx1 sin(δ). (9)

Fig. 7, Fig. 8, Table 1 and Table 2 present SMO observer results for the
two tests. These tables show the mean and the standard deviation of the
normalized errors. The state estimates were initialized with the maximum
measurements during the test (for instance xb2 (t = 0) = 5448N in the slalom

7
 Front lateral tire force Fyw1 (kN)
5
5448 N

0
SMO
-5 Measurement
0 5 10 15 20 25

Rear lateral tire force F (kN)

yw2

2
0
SMO
-2
Measurement
0 5 10 15 20 25

Front longitudinal tire force Fxw1 (kN)

5

SMO
-5 Measurement
0 5 10 15 20 Time (s)

Fig. 7. Slalom test. SMO observer. Tire force estimates.

Table 1
Slalom test, SMO observer, maximum absolute values, normalized mean errors and
standard deviation (Std)
Slalom max |zmeas | mean(εz ) std(εz )
Fy1 6282 N 4.1% 3.8%
Fy2 3663 N 2.2% 1.9%
Fx1 6181 N 3.8% 3.6%
ψ̇ 33.6 o /s 0.5% 0.3%

test). In spite of these initializations the estimates converge quickly to the

measurements, showing the good convergence properties of the observer.
Moreover, the SMO observer produces satisfactory estimates close to mea-
surements for the two tests (normalized mean errors and standard deviation
less than 6 % in Table 1 and Table 2). These good experimental results con-

8
 Front lateral tire force F (kN)
yw1
4

0
SMO
-2 Measurement
0 10 20 30 40 50

Rear lateral tire force F (kN)

yw2

0
SMO
Measurement
-2
0 10 20 30 40 50

Front longitudinal tire force F (kN)

xw1

-2 SMO
Measurement
0 10 20 30 40 Time (s)

Fig. 8. Roundabout. SMO observer. Tire force estimates.

firm that the observer approach may be appropriate for the estimation of tire
forces.

Table 2
Roundabout test, SMO observer, maximum absolute values, normalized mean errors
and standard deviation (Std)
Roundabout max |zmeas | mean(εz ) std(εz )
Fy1 4010 N 5.2% 3.4%
Fy2 3215 N 3.7% 3.2%
Fx1 3434 N 4.8% 3.9%
ψ̇ 23.0 o /s 0.2% 0.2%

9
3 Sideslip angle and wheel cornering stiffness observer, EKF

This section introduces tire-force modeling and presents the observer EKF for
the estimation of sideslip angle and wheel cornering stiffness. The filter EKF
is constructed by combining models characterizing the vehicle sideslip angle
and the tire-road contact forces. The sideslip angle evolution is formulated by
using the single-track model, (Segel, 1956):

1 (10)
β̇ = mVg
[Fxw1 sin(δ − β) + Fyw1 cos(δ − β) + Fyw2 cos(β)] − ψ̇.

The wheel sideslip angles and the center of gravity sideslip angle are linked
with the following relationships:

L1 ψ̇ L2 ψ̇ (11)
β1 = δ − β − Vg
, β2 = −β + Vg
.

The tire forces are usually modeled as a function of slips between tire and
road, such as the wheel longitudinal slip and the wheel sideslip angle (Pacejka
& Bakker, 1991; Burckhardt, 1993; Canudas-De-Wit, Tsiotras, Velenis, Basset
& Gissinger, 2003). Fig. 9 illustrates different lateral tire-force models (linear,
linear adaptive and Burckhardt for various road surfaces, (Kiencke & Nielsen,
2000)).
For normal driving situations, lateral tire forces are usually considered linear

5
Ci
4.5

4
Lateral tire force (kN)

3.5

2.5

2 Ci+∆Cai Linear
Linear adaptive
1.5
Burckhardt, asphalt dry
1 Burckhardt, asphalt wet
Burckhardt, cobblestone wet
0.5 Burckhardt, ice
0
0 2 4 6 8 10 12 14 16
Wheel sideslip angle (°)

Fig. 9. Lateral tire force models: linear, linear adaptive, Burckhardt for various road
surfaces.

10
with respect to sideslip angle (linear model, cumulative lateral tire-force for
each axle):

Fywi (βi ) = Ciβi , i = 1, 2, (12)

where Ci is the wheel cornering stiffness (a parameter closely related to tire-

road friction).
When road friction changes or when the nonlinear tire domain is reached,
wheel cornering stiffness varies. Some studies have described observers which
take cornering stiffness (or road friction) variations into account (Rabhi, M’Sirdi,
Zbiri & Delanne, 2005; Ray, 1997). In (Rabhi, M’Sirdi, Zbiri & Delanne, 2005)
tire-road parameters are identified with a sliding-mode observer, while in (Ray,
1997) tire-road forces are modeled with an integrated random walk model
(F̈ = 0).
In this study, an adaptive tire-force model is proposed (hereafter termed the
”linear adaptive force model”). This model is based on the linear model to
which a readjustment variable ∆Cai has been added to correct wheel corner-
ing stiffness errors (Fig. 9):

Fywi (βi ) = (Ci + ∆Cai )βi , i = 1, 2. (13)

The variable ∆Cai is included in the state vector of the EKF observer and its
evolution equation is formulated according to the model (∆Ċai = 0).
The EKF observer is constructed with state X′ ∈R3 , input U′ ∈R4 and mea-
surement Y ′∈R3 as:

X′ = [x′1 , x′2 , x′3 ] = [β, ∆Ca1 , ∆Ca2 ],

U′ = [u′1 , u′2 , u′3 , u′4] = [δ, ψ̇, Vg , Fxw1], (14)
Y ′ = [y1′ , y2′ , y3′ ] = [Fyw1 , Fyw2 , γy ].

The measurement model is

y1′ = (C1 + x′2 )β1 , y2′ = (C2 + x′3 )β2 ,

(15)
1
y3′ = m
[(C1 + x′2 )β1 cos(u′1 ) + (C2 + x′3 )β2 + u′4 sin(u′1 )].

Consider the state estimates denoted as xb′ = [xb′1 , xb′2 , xb′3 ]. The state evolution
model of EKF is:
′ u′4 sin(u′1 − xb′1 ) + Fauxi1 cos(u′1 − xb′1 ) + Fauxi2 cos(xb′1 )
xḃ1 = − u′2 ,
mu′3 (16)
′ ′
xḃ2 = 0, xḃ3 = 0,

11
where the auxiliary variables Fauxi1 and Fauxi2 are calculated as:

L1 u′2
Fauxi1 = (C1 + xb′2 )(u′1 − xb′1 − u′3
),
(17)
L2 u′2
Fauxi2 = (C2 + xb′3 )(−xb′1 + u′3
).

The extended Kalman filter EKF is based on a study by (Kalman, 1960)

and constructed according to an algorithm proposed in (Mohinder & Angus,
1993). The EKF algorithm is set according to noise variance-covariance ma-
trices. These matrices, denoted as Qs , Qe and Qm , are associated respectively
to the noises of the measurements, the inputs and the models. These noises
are assumed to be Gaussian, white and centered. The variance-covariance ma-
trices are set according to the vehicle state. This is explained with the Figs.
10 and 11.
Fig. 10 presents the measurements of the front lateral tire forces Fy1 as a func-

4
Front lateral tire force Fy1 (kN)

-2

-4

-6

-8
-5 -4 -3 -2 -1 0 1 2 3 4
Front sideslip angle β1 (°)

Fig. 10. Experimental test measurements of the front lateral tire forces as a function
of the front wheel sideslip angles.

tion of the front wheel sideslip angles β1 (for a slalom test). Fig. 11 illustrates
straight lines representing different potential cornering stiffnesses estimates
(Ci + ∆Cai = Fyi /βi ). Neglecting the force saturation, the estimation of cor-
nering stiffness is not adapted in four zones:

• in zones 1 and 2: when βi ∗ Fyi < 0 (not adapted because the cornering
stiffness must be positive),

12
 6
Zone 5

Zone 1
λ1
Front lateral tire force Fy1 (kN)
2
Zone 3
0
Zone 4
-2
− λ1

Zone 2
-4
C1+∆Ca1

-6
Zone 6
-8
-5 -4 -3 -2 -1 0 1 2 3 4
Front sideslip angle β1 (°)

Fig. 11. Experimental test measurements of the front lateral tire forces as a function
of the front wheel sideslip angles. The straight lines represent different cornering
stiffnesses (Ci + ∆Cai = Fyij /βij ). State zones 1, 2, 3 and 4 are not adapted for
the estimation of cornering stiffness. The observer EKF mainly corrects cornering
stiffness in zones 5 and 6.

• in zones 3 and 4: when the lateral dynamic is low: Fyi < λi (threshold),
because some ratios Fyij /βij may have very different values in these zones.

The variance-covariance matrices (Qs , Qm ) are set in order that the cornering
stiffness estimation is performed mainly in zones 5 and 6.
The variance-covariance matrices are set

• in zone 1,2,3,4:
Qs = diag[1e6,1e6,1e6],
Qm = diag[1e-13, 0, 0],

• in zone 5,6:
Qs = diag[0.1, 0.1, 0.1],
Qm = diag[1e-13, 0.24, 0.24].

An observability function was calculated using a Lie derivative method, (Ni-

jmeijer & Van der Schaft, 1990). The rank of the observability function, which
was calculated along experimental trajectories, corresponded to the state vec-
tor dimension (3), and consequently system EKF was locally observable.

13
3.1 Experimental results - observer EKF

In order to demonstrate the improvement provided by the observer using the

linear adaptive force model (13), another observer constructed with a linear
force model (12) was used in comparison (observer denoted ORL). The ro-
bustness of the two observers was tested with respect to tire-road friction
variations by performing the tests with different cornering stiffness parame-
ters ([C1 , C2 ] ∗ 0.5, 1, 1.5 with C1 = 65000 N.rad−1 and C2 = 50000 N.rad−1 ).
The different observers were evaluated for the two tests presented in section
2.2.
Figs. 12-13 show the results for the ORL observer (using the linear force model)
when estimating rear sideslip angle. The ORL observer gives good results
Slalom, ORL (linear model), rear sideslip angle (°)
measurement
5 (C , C )*1.5
1 2

0
(C , C )
1 2

-5 (C , C )*0.5
1 2

0 5 10 15 20 Time (s)

Fig. 12. ORL observer (linear tire-force model), rear sideslip angle estimates,
with different stiffness settings.

Roundabout, ORL (linear model), rear sideslip angle (°)

measurement
4
(C , C )*1.5
1 2
2
(C , C )
1 2
0
(C 1 , C 2 )*0.5
-2
-4
-6
0 10 20 30 40 Time (s)

Fig. 13. ORL observer (linear tire-force model), rear sideslip angle estimates,
with different stiffness settings.

when cornering stiffnesses are approximately known ([C1 , C2 ] ∗ 1). However,

this observer is not robust when cornering stiffnesses change ([C1 , C2 ]∗0.5, 1.5).
When cornering stiffness is reduced ORL overestimates rear sideslip angle
(in absolute value) and, conversely, when cornering stiffness is increased rear
sideslip angle is overestimated.
Fig. 14, Fig. 15, Table 3 and Table 4 present rear sideslip angle estimation

14
results obtained with the adaptive observer EKF for the two tests. The per-
formance robustness of EKF is satisfactory since rear sideslip angle is well
estimated and similar irrespective of cornering stiffness settings. This result
is confirmed by the EKF normalized errors (Table 3 and Table 4) which are
approximately similar.

Slalom, EKF (adaptive model), rear sideslip angle (°)

5 measurement
(C 1 , C 2 )*1.5

0
(C , C )
1 2
(C 1 , C 2 )*0.5
-5
0 5 10 15 20 Time (s)

Fig. 14. Slalom test, EKF observer (linear adaptive force model), with dif-
ferent cornering stiffness settings Rear sideslip angle estimates.

Table 3
Slalom test, EKF observer, maximum absolute values, normalized mean errors and
standard deviation (Std) of the rear sideslip angle
Slalom (C1 , C2 ) ∗ 0.5 (C1 , C2 ) (C1 , C2 ) ∗ 1.5
max |β2 | 5.7o 5.7o 5.7o
mean(εz ) 5.0% 4.4% 5.1%
std(εz ) 5.9% 5.2% 5.5%

Roundabout, EKF (adaptive model), rear sideslip angle (°)

5 measurement
(C , C )*1.5
1 2
(C , C )
1 2
0
(C , C )*0.5
1 2

-5
0 10 20 30 40 Time (s)

Fig. 15. Slalom test, EKF observer (linear adaptive force model), with dif-
ferent cornering stiffness settings. Rear sideslip angle estimates.

Fig. 16 and Fig. 17 present the front and rear cornering stiffness estimates
(Ci + ∆Ci ). It can be seen that the cornering stiffness estimates remain ap-

15
Table 4
Roundabout test, EKF observer, maximum absolute values, normalized mean errors
and standard deviation (Std) of the rear sideslip angle
Roundabout (C1 , C2 ) ∗ 0.5 (C1 , C2 ) (C1 , C2 ) ∗ 1.5
max |β2 | 5.4o 5.4o 5.4o
mean(εz ) 8.1% 6.4% 5.9%
std(εz ) 7.4% 6.6% 6.5%

proximately constant when the lateral dynamic (sideslip angles, lateral tire
forces) is low, that is to say during the first 10 s in the case of the slalom,
and the first 9 s in the case of the roundabout. Then, at the beginning of the
slalom at 10 s or at the beginning of the roundabout at 9 s, the lateral dy-
namic becomes sufficiently large, and the estimates of the cornering stiffness
converge quickly to the same values. The new information obtained from ∆Cai
may be used as an indication of the road friction. Indeed, for same values of
the wheel sideslip angle, a significative change of ∆Cai variable can indicate a
road friction change (Baffet (2007)).

4 Slalom, EKF, Front cornering stiffness (N.rad -1 )

x 10
(C 1 , C 2)*1.5
8 (C 1 , C 2 )
(C 1 , C 2 )*0.5
6

4
0 5 10 15 20 25

-1
4 Slalom, EKF, Rear cornering stiffness (N.rad )
x 10
8 (C , C )*1.5
1 2
(C , C )
1 2
6 (C , C )*0.5
1 2

0 5 10 15 20 Time (s)

Fig. 16. Slalom test, EKF observer (linear adaptive force model), with dif-
ferent cornering stiffness settings [(C1 , C2 ) = (65000, 50000) ∗ (0.5, 1, 1.5) N.rad−1 ].
Front and rear cornering stiffness estimates Ci + ∆Cai .

16
 -1
4 Roundabout, EKF, Front cornering stiffness (N.rad )
x 10
(C , C )*1.5
1 2
8 (C , C )
1 2

6 (C , C )*0.5
1 2

0 10 20 30 40 50

-1
4 Roundabout, EKF, Rear cornering stiffness (N.rad )
x 10
8
(C 1 , C 2 )*1.5
(C , C )
6 1 2
(C , C )*0.5
1 2

0 10 20 30 40 Time (s)

Fig. 17. Roundabout test, EKF observer (linear adaptive force model),
with different cornering stiffness settings [(C1 , C2 ) = (65000, 50000) ∗ (0.5, 1, 1.5)
N.rad−1 ]. Front and rear cornering stiffness estimations Ci + ∆Cai .
4 Sensitivity to vehicle mass

This section presents the sensitivity of the estimation process when the vehicle
mass is unknown. Three tests are performed with a different mass (m = 1447
kg, m+300, m−300). Fig. 18 and Table 5 represent the results. Concerning the
tire forces, estimates raise when the mass parameter is increased (in absolute
value), and conversely, estimates decrease when the mass is reduced. The tire
force estimates converge to different values, showing that SMO is sensitive to
changes in vehicle mass. As regard rear sideslip angle, the variations induced by
mass change are lower than those of the tire force estimates. This is confirmed
by the Table 5, where normalized mean errors of Fy2 varies from 2.2 % to 10.4
%, whereas normalized mean errors of β2 changes from 4.4 % to 5.4 %. This
can be explained by regarding the sideslip angle model (10), where the mass
change in the denominator is partially compensated for the force change in
the numerator.

5 Conclusion

This study deals with two vehicle-dynamic observers constructed for use in a
two-block estimation process. The block 1 mainly estimates tire-forces (with-

17
Table 5
Slalom test, robustness to vehicle mass (m = 1447 kg). Maximum absolute values,
normalized mean errors and standard deviation (Std)
Slalom m − 300 kg m m + 300 kg
Fy1 max|zmeas | 6282 N 6282 N 6282 N
mean(ǫz ) 8.0% 4.1% 4.9%
std(ǫz ) 8.2% 3.8% 4.8%
Fy2 max|zmeas | 3663 N 3663 N 3663 N
mean(ǫz ) 10.4% 2.2% 4.7%
std(ǫz ) 10.1% 1.9% 4.5%
Fx1 max|zmeas | 6181 N 6181 N 6181 N
mean(ǫz ) 5.9% 3.8% 5.5%
std(ǫz ) 5.9% 3.6% 8.1%
β2 max|zmeas | 5.7o 5.7o 5.7o
mean(ǫz ) 5.4% 4.4% 5.1%
std(ǫz ) 5.8% 5.2% 5.7%

SMO, Front longitudinal tire force Fxw1 (kN) SMO, Rear lateral tire force Fyw2 (kN)
5 m-300 m-300
m m
m+300 2 m+300

0 0

-2
-5 measurement
measurement
-4
0 5 10 15 20 25 0 5 10 15 20 25
SMO, Front lateral tire force Fyw1 (kN) EKF, Rear sideslip angle (°)
5 5 m-300
m
m+300
0
0
m-300
-5 m
m+300
measurement measurement
-5
0 5 10 15 20 25 0 5 10 15 20 25

Fig. 18. Slalom test, with different mass settings: m + 300 kg, m = 1447 kg, m − 300
kg.

out an explicit tire-force model), while the block 2 calculates sideslip angle
and corrects cornering stiffnesses (with an adaptive tire-force model).
The first observer SMO (block 1), a sliding mode observer, was constructed
by finding the gains with respect to the convergence. The experimental evalu-
ations of SMO are satisfactory, showing estimates close to the measurements

18
and good convergence properties.
The second observer EKF (block 2), an extended Kalman filter combined with
an adaptive tire-force model, was evaluated for different cornering stiffness set-
tings and was compared with an observer constructed with a fixed tire-force
model (ORL). Results show that ORL is not robust when cornering stiffness
parameters change, whereas EKF gives good estimates of the sideslip angle.
This result justifies the use of an adaptive tire-force model.
The different results show the potential of the two-block estimation process.
The first block has the advantage of providing satisfactory force estimates
without a tire-force model, whereas the second block provides robust sideslip
angle estimates with respect to cornering stiffness changes.

Acknowledgements

This work was supported by the PREDIT-SARI-RADARR program. The au-

thors express gratefulness for the reviewer inputs that improve readability and
proof of the sliding mode observer.

A Convergence of the SMO observer

f
The aim is to set gains Ω1 , ..., Ω12 which lead to the state estimation errors X
converging toward 0. The evolution equations for the state estimation errors
are the following:

ḟ = Ẋ − x,
X ḃ
1
xė1 = Iz
[L1 xe2 − L2 xe3 ] − Ω1 sign(ye1 ) − Ω2 sign(ye2 ) − Ω3 sign(ye3),
xė2 = −Ω4 sign(ye1 ) − Ω5 sign(ye2) − Ω6 sign(ye3 ), (A.1)
xė3 = −Ω7 sign(ye1 ) − Ω8 sign(ye2) − Ω9 sign(ye3 ),
xė4 = −Ω10 sign(ye1) − Ω11 sign(ye2 ) − Ω12 sign(ye3).

Proposition 1. Consider the following Lyapunov functions (Φ1 , Φ2 , Φ3 , Φ4 ):

Φ1 = 21 xe21 , Φ2 = 12 xe24 , Φ3 = 21 (xe2 + xe3 )2 ,

(A.2)
Φ4 = 12 (L1 xe2 − L2 xe3 )2 .

If the derivatives of the Lyapunov functions are negative, then the estimates
b will converge toward the system states X.
x

19
Proof 1. The Lyapunov functions Φ1 , ..., Φ4 are positive along all state tra-
jectories, and consequently if the derivatives Φ̇1 , ..., Φ̇4 are negative, then the
Lyapunov functions will converge toward zero. This entails the following con-
vergence:

xe1 −→ 0, xe4 −→ 0, xe2 + xe3 −→ 0, L1 xe2 − L2 xe3 −→ 0. (A.3)

and consequently: xb1 −→ x1 , xb4 −→ x4 . The convergence xe2 + xe3 −→ 0 implies

two possibilities at the switching surface:

• either [xe2 −→ 0 and xe3 −→ 0], then xb2 −→ x2 and xb3 −→ x3 ,

• or [xe2 −→ −xe3 ], then the convergence A.3 implies xe2 (L1 + L2 ) −→ 0, this
entails the convergence xb2 −→ x2 and xb3 −→ x3 .

Remark 1. The Lyapunov functions A.2 correspond to the estimation errors

for: yaw rate ψ̇ (Φ1 ), lateral acceleration γy (Φ2 ), yaw acceleration ψ̈ (Φ3 ),
longitudinal acceleration γx (Φ4 ). Gains Ω1 , ..., Ω12 are chosen so as to make
the Lyapunov function derivatives negative.

Proposition 2. If the gains are selected such that:

Ω1 ≥ |Ω2 | + |Ω3 | + | I1z (L1 x̃2 + L2 x̃3 )| + ǫ,

(A.4)
Ω12 ≥ |Ω11 | + |Ω10 | + ǫ,

where ǫ > 0 then the estimates of the yaw rate x̂1 and longitudinal tire force
x̂4 converge toward the states x1 and x4 respectively.

Proof 2a. The derivative of the Lyapunov function Φ1 is:

Φ̇1 = x̃1 x̃˙ 1 ,

Φ̇1 = x̃1 [ I1z (L1 x̃2 − L2 x̃3 ) − Ω1 sign(ỹ1 ) − Ω2 sign(ỹ2)
− Ω3 sign(ỹ3 )], (A.5)
Φ̇1 = −x̃1 [Ω1 sign(x̃1 ) + Ω2 sign(x̃2 + x̃3 ) + Ω3 sign(x̃4 )
1
− Iz
(L1 x̃2 − L2 x̃3 )].

Using the triangle inequality and A.4 now gives the following inequality for
the Lyapunov function

Φ̇1 ≤ −x̃1 ǫ sign(x̃1 ) < 0.

20
This derivative is negative and Φ1 is positive, and consequently this induces
the convergence Φ1 −→ 0 x̃1 −→ 0 and therefore x̂1 −→ x1 .

Remark 2a. At the switching surface xė1 = 0 this result implies:

1 (A.6)
signeq (xe1 ) = Ω1 I z
[L1 xe2 − L2 xe3 ].

Proof 2b. The derivative of the Lyapunov function Φ2 is:

Φ̇2 = x̃4 x̃˙ 4 ,

= x̃4 [−Ω10 sign(ỹ1 ) − Ω11 sign(ỹ2) − Ω12 sign(ỹ3 )] , (A.7)
= x̃4 [−Ω10 sign(x̃1 ) − Ω11 sign(x̃2 + x̃3 ) − Ω12 sign(x̃4 )] .

Using the triangle inequality and A.4 now gives the following inequality on
the Lyapunov function Φ2

Φ̇2 ≤ −x̃4 ǫ sign(x̃4 ) < 0.

This derivative is negative and Φ2 is positive, and consequently this implies

the convergence Φ2 −→ 0, x̃4 −→ 0 and therefore x̂4 −→ x4 .

Proposition 3. Choosing the gains such that:

Ω5 + Ω8 ≥ |Ω6 | + |Ω9 | + ǫ,
Ω4 = −Ω7 > 0, (A.8)
Ω5 = Ω8 L2 /L1 ,

where ǫ > 0, implies convergence of lateral tire force estimates (x̂2 , x̂3 ) toward
the states (x2 , x3 ).

Proof 3a. The derivative of the Lyapunov function Φ3 is:

Φ̇3 = (x̃2 + x̃3 )(x̃˙ 2 + x̃˙ 3 ),

= (x̃2 + x̃3 )[−(Ω4 + Ω7 )sign(ỹ1) − (Ω5 + Ω8 )sign(ỹ2 ) (A.9)
− (Ω6 + Ω9 )sign(ỹ3 )].

Using A.8 and the triangle inequality yields the following inequality:

Φ̇3 ≤ −(x̃2 + x̃3 ) ǫsign(x̃2 + x̃3 ) < 0.

21
Consequently, with Φ3 positive this implies convergences Φ3 −→ 0 and x̃2 +
x̃3 −→ 0.

Proof 3b. The derivative calculation of Lyapunov function Φ4 gives:

Φ̇4 = (L1 xe2 − L2 xe3 )(L1 xė2 − L2 xė3 ),

Φ̇4 = (L1 xe2 − L2 xe3 )[(−L1 Ω4 + L2 Ω7 )sign(ye1 ) (A.10)
+ (−L1 Ω5 + L2 Ω8 )sign(ye2 )].

Applying the result (A.6) and choosing Ω4 = −Ω7 , (A.10) becomes:

Φ̇4 = (L1 xe2 − L2 xe3 )[ ΩΩ17Iz (L1 + L2 )(L1 xe2 − L2 xe3 )

(A.11)
+ (−L1 Ω5 + L2 Ω8 )sign(ye2 )].

Specifying Ω5 = Ω8 L2 /L1 and Ω7 < 0 means that the derivative:

Ω7 (L1 +L2 )
Φ̇4 = Ω1 I z
(L1 xe2 − L2 xe3 )2 < 0, (A.12)

becomes negative, which implies the convergences Φ4 −→ 0, and therefore

L1 xe2 − L2 xe3 −→ 0. Moreover, xe2 + xe3 −→ 0, consequently this induces the
convergence (xb2 , xb3 ) −→ (x2 , x3 ).

B Nomenclature

Table B.1 and Table B.2 present the different notations.

References

Arndt, C., & Karidas, J., 2006, Estimation of tire-road contact for integrated
control applications in current vehicles. Proceedings of the IEEE Conference
on Control Application, CCA, Munich, Germany, 2784–2789.
Baffet, G., Stephant, J., & Charara, A., 2007, Experimental evaluation of tire-
road forces and sideslip angle observers. Proceedings of the IEEE European
Control Conference, ECC, Kos, Greece.
Baffet, G., 2007, Developpement et validation experimentale d observateurs
des forces du contact pneumatique/chaussee d une automobile, Ph.D. dis-
sertation, Universite de Technologie de Compiegne, Compiegne, France.

22
Table B.1
Nomenclature (first part).
Symbol Description
C1 Front cornering stiffness (N.rad−1 )
C2 Rear cornering stiffness (N.rad−1 )
Fx1 Front longitudinal tire force (in body car axis) (N )
Fy1 Front lateral tire force (in body car axis) (N )
Fxw1 Front longitudinal tire force (in wheel axis) (N )
Fxw2 Rear longitudinal tire force (in wheel axis) (N )
Fyw1 Front lateral tire force (in wheel axis) (N )
Fyw2 Rear lateral tire force (in wheel axis) (N )
Ω1 , ..., Ω20 SMO sliding mode observer gains
Iz Yaw moment of inertia (kg.m−2 )
L1 Center of gravity to front axle distance (m)
L2 Center of gravity to rear axle distance (m)
m Vehicle mass (kg)
Qe Variance-covariance matrix for input noises
Qm Variance-covariance matrix for model noises
Qs Variance-covariance matrix for measurement noises
Φ1 , ..., Φ5 Lyapunov functions
Vg Center of gravity vehicle speed (m.s−1 )
Vg1 Front vehicle speed (m.s−1 )
Vg2 Rear vehicle speed (m.s−1 )
β Center of gravity sideslip angle (rad)
β1 Front sideslip angle (rad)
β2 Rear sideslip angle (rad)
δ Front mean steering angle (rad)
∆Ca1 Front adaptive cornering stiffness (N.rad−1 )
∆Ca2 Rear adaptive cornering stiffness (N.rad−1 )
ψ̇ Yaw rate (rad.s−1 )
γx Longitudinal accelerations (m.s−2 )
γy Lateral accelerations (m.s−2 )

23
Table B.2
Nomenclature (second part).
Symbol Description
x, y Longitudinal and lateral vehicle positions (m)
X, Y SMO, state, measurement vectors
b
x SMO, state estimation vector
e
X SMO, vector of state estimation errors
b
y SMO, vector of measurement estimates
e
Y SMO, vector of measurement estimation errors
X′ , U′ , Y ′ EKF, state, input, measurement vectors
b′
x EKF, state estimation vector
b′
y EKF, vector of measurement estimates

Burckhardt, M., 1993, Fahrwerktechnik: Radschlupfregelsysteme. Vogel-

Verlag.
Canudas-De-Wit, C., Tsiotras, P., Velenis, E., Basset, M., & Gissinger, G.,
2003, Dynamic friction models for road/tire longitudinal interaction. Vehicle
System Dynamics, vol. 39, pp 189-226.
Drakunov, S., & Utkin, V., 1995, Sliding mode observers. IEEE Conference
on Decision and Control, pp 3376-3379.
Gustafsson, F., 1997, Slip-based Tire-Road Friction Estimation. Automatica,
vol. 33, no 6, pp. 1087-1099.
Hsu, Y.H.J, & Gerdes, J.C., 2006, A feel for the road: a method to estimate
tire parameters using steering torque. Proceedings of the 8th International
Symposium on Advanced Vehicle Control, AVEC, Taipei, Taiwan, 835–840.
Kalman, R.E., 1960, A New Approach to Linear Filtering and Prediction
Problems. Transactions of the ASME - Journal of Basic Engineering, vol.
82, série D, pp. 35-45.
Kiencke, U., & Nielsen, L., 2000, Automotive control system. Springer.
Kiencke, U., & Daib, A., 1997, Observation of lateral vehicle dynamics. Control
Engineering Practice, vol. 5, Issue 8, pp. 1145-1150.
Kim. J., 2009, Identification of lateral tyre force dynamics using an extended
Kalman filter from experimental road test data. Control Engineering Prac-
tice, vol. 17, Issue 3, March 2009, pp. 357-367.
Lakehal-ayat, M., Tseng, H.E., Mao, Y., & Karidas, j., 2006, Disturbance Ob-
server for Lateral Velocity Estimation. Proceedings of the 8th International
Symposium on Advanced Vehicle Control AVEC, Taipei, Taiwan, 889–894.
Lechner, D., 2002, Analyse du comportement dynamique des vehicules routiers
legers: developpement d’une methodologie appliquee a la securite primaire.
Ph.D. dissertation, Ecole Centrale de Lyon, Lyon, France.
Levant, A., 1993, Sliding order and sliding accuracy in sliding mode control,

24
 Int. Journal of Control, Vol. 58, no. 6, pp. 1247-1263.
Mohinder, S.G., & Angus, P.A., 1993, Kalman filtering theory and practice.
Prentice hall.
Nijmeijer, H., & Van der Schaft, A.J., 1990, Nonlinear Dynamical Control
Systems. Springer-Verlag.
Ono, E., Asano, K., Sugai, M., Ito, S., Yamamoto, M., Sawada, M., & Yasui,
Y., 2003, Estimation of automotive tire force characteristics using wheel
velocity. Control Engineering Practice, Vol. 11, Issue 12, pp. 1361-137.
Pacejka, H.B., & Bakker, E., 1991, The magic formula tyre model. Proceedings
of 1st Int. colloq. on tyre models for vehicle dynamics analysis, pp. 1-18.
Rabhi, A., M’Sirdi, N.K., Zbiri, N., & Delanne, Y., 2005, Vehicle-road interac-
tion modelling for estimation of contact forces. Vehicle System Dynamics,
Taylor and Francis, Vol. 43, Supplement, pp. 403-411.
Ray, L., 1997, Nonlinear Tire Force Estimation and Road Friction Identifi-
cation: Simulation and Experiments. Automatica, vol. 33, no. 10, pp 1819-
1833.
Segel, M.L., 1956, Theorical prediction and experimental substantiation of the
response of the automobile to steering control. Proceedings of automobile
division of the institut of mechanical engineers, vol. 7, pp 310-330.
Slotine, J., Hedrick, J., & Misawa, E., 1987, On sliding observer for nonlinear
systems. Journal of mathematical systems. Estimation and control, vol. 109,
pp 245-259.
Stephant, J., Charara, A., & Meizel, D., 2007, Evaluation of a sliding mode
observer for vehicle sideslip angle. Control Engineering Practice, vol. 15, pp
803–812.
Tanelli, M., Savaresi, S., & Cantoni, C., 2006, Non-local Extremum seeking
control for active braking control systems. Proceedings of the IEEE Confer-
ence on Control Application, CCA, Munich, Germany, 891–896.
Ungoren, A.Y., Peng, H., & Tseng, H.E., 2004, A study on lateral speed esti-
mation methods. Int. J. Vehicle Autonomous Systems, vol. 2, Nos. 1/2, pp.
126-144.
Wesemeier D., Isermann R., Identification of vehicle parameters us-
ing stationary driving maneuvers. Control Engineering Practice,
vdoi:10.1016/j.conengprac.2008.10.008.

25