2008 American Control Conference

Westin Seattle Hotel, Seattle, Washington, USA

June 11-13, 2008

FrB09.3

Estimation of the maximum friction coefficient for a passenger vehicle

using the instantaneous cornering stiffness

Abstract The estimation of the maximum lateral coefficient

of friction of a passenger vehicle between tire and road is
presented in this paper. This is achieved by utilizing the
instantaneous cornering stiffness, which is defined as the slope
of the nonlinear curve of the lateral friction coefficient at the
instantaneous tire slip angle. The maximum lateral coefficient of
friction is necessary for integrated global cassis control (IGCC)
especially for the estimation of the lateral velocity of the vehicle.
The advantage of this method is the low computational effort
and the independence of tire or road conditions. Two methods
for estimation of the instantaneous cornering stiffness and the
maximum coefficient of friction in real time and a method
for offline estimation used as a reference are described and
validated using measured data of a passenger vehicle.

friction coefficient y

Lukas Haffner, Martin Kozek, Jingxin Shi, and H. Peter Jorgl

C () = 0

0.8
0.6

y,max

0.4

C (0)

0.2
0
0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

tire slip angle

Fig. 1. Lateral tire model presenting the instantaneous cornering stiffness
C ()

I. INTRODUCTION
For IGCC it is necessary to have precise information
about vehicle states and the tire-road condition. A main
problem is to detect when the tire force reaches its maximum
and the tires start to skid. This may result in a unstable
driving situation. Many works have already been carried
out on the slip based estimation of the maximum friction
coefficient max between tire and road. The longitudinal
slip and the slip angle are used for longitudinal and
lateral dynamics, respectively. In the following publications
a defined slope is used to identify max for longitudinal and
lateral considerations:
In [1] the slip slope k is defined as the slope of the x -curve at x = 0. The information of the slip slope and the
variance of the slip signal leads to a defined road condition
including a specific x,max . With this method four different
road surfaces can be classified. In [2] the same assumption
is made distinguishing between dry and wet asphalt only by
estimating the slip slope. In [3] the extended braking stiffness
XBS is defined as the slope of the Fx --curve at any
with Fx being the longitudinal force of the tire. The XBS
is estimated by the power spectrum density of the angular
velocities of the wheels.
The cornering stiffness c0 is estimated in [4] to be used in
active steering. c0 is defined as the slope of the Fy --curve
at = 0. In [5] the cornering stiffness C is estimated using
the same definition like [4].
This paper concentrates on lateral dynamics including lateral
L. Haffner, M. Kozek, and H. P. Jorgl are with the Institute of Mechanic
and Mechatronic, Division of Control and Process Automation, Vienna
University of Technology, A-1040 Vienna, Austria

lukas.haffner@tuwien.ac.at
martin.kozek@tuwien.ac.at
J. Shi is with TTTech Germany GmbH, D-85276 HettenshausenReisgang, Germany

jingxin.shi@tttech.de

978-1-4244-2079-7/08/$25.00 2008 AACC.

tire forces and tire slip angles. In contrast to [4] and [5] the
instantaneous cornering stiffness C () is defined as first
derivative of the y --curve similar to the definition in [3]
for longitudinal dynamics. Thus it is not constant for static
road conditions, but changes regarding the slope of the y -curve.
For C () = 0, the maximum lateral coefficient of friction
y,max is reached. Hence, the vehicle state has to pass the
non-monotone part of the lateral tire characteristics in order
to detect y,max as shown in Fig. 1 where C () becomes
zero or negative. An actual vehicle in such a driving situation
will become unstable, and therefore the estimation of y,max
is essential for IGCC.
The proposed concept of instantaneous cornering stiffness
estimation is therefore capable of not only detecting y,max ,
but also of estimating the instantaneous driving conditions.
The remainder of the paper is organized as follows: In II
a vehicle model and the used variables are defined, in III
three estimation methods for C () are presented, which is
consequently utilized as an input to the y,max -estimation,
in IV experimental results show the performance of these
methods, and in V conclusions and future works are shown.
II. VEHICLE MODEL
The single-track model of Fig. 2 is used for further
and
considerations with vCoG , ax , ay , Fy , v, , l, ,
being the velocity of the center of gravity, the longitudinal
and lateral acceleration, the lateral force, the velocity of the
tire, the tire slip angle, the distance from center of gravity
to the tire, the yaw rate, the steering angle, and the sideslip
angle of the vehicle, respectively. The indices F and R relate
to the front and rear tire of the vehicle, respectively. The road
inclination and superelevation are not taken into account.
However, there exists literature where the problem of vehicle

4591

III. ESTIMATION OF THE INSTANTANEOUS

CORNERING STIFFNESS AND THE FRICTION
COEFFICIENT

vF
lF

A. Basic Considerations

vCoG
Fy,F

y,max

vR

ay

ax
ay

Fy,R
Single-track model for vehicle lateral dynamics

state estimation in the presence of these road characteristics

is treated extensively, e.g. [6].
The lateral and vertical tire forces Fy and Fz can be written
as
Fy,F

Fy,R

Fz,F

Fz,R

may lR Jz
,
(lF + lR )cos
may lF + Jz
,
lF + lR
m(glF ax h)
, and
lF + lR
m(glR + ax h)
lF + lR

(1)

(2)

with m, Jz and g being the vehicle mass, the moment of

inertia and the gravitatonal constant. The front and rear
lateral coefficient of friction read:
y,F

y,R

Fy,F
Fz,F
Fy,R
Fz,R

(3)

y is a nonlinear function depending on the tire slip angle,

the vehicle vehicle velocity, the tire and road condition
and other parameters. The slip angles of the front and
rear tires can be calculated with the following widely used
approximation:
F

F
l
vCoG

lR
+
vCoG

(4)

Using the relations (1) to (4) the lateral vehicle dynamics can
be computed for all driving situations. Since not all necessary
variables are measured online, a combination of parameter
estimation and observers becomes necessary (see Fig. 3). In
order to reconstruct correct tire slip angles F and R in
all driving conditions the cornering stiffness is of crucial
importance.

F
-observer

F -calculation

Fig. 2.

y,max -estimation

Fig. 3.

Fy,F
Fy,R
Fz,F
Fz,R

y,F
y,R

C ()-estimation

-calculation

lR

C,F
C,R

Schematic of vehicle state estimation

For the estimation of the instantaneous cornering stiffness

C () the inputs y and are required. y can be calculated
with (1), (2) and (3) as ax , ay and are measured in
modern passenger vehicles. has to be either measured or
estimated by an observer. The measurement of can be
done indirectly by velocity sensors in longitudinal and lateral
direction and with (4). Velocity sensors are expensive and
not feasible in mass production. Alternatively, an -observer
uses ax , ay , , and the four angular velocities of the
wheels ~
as inputs. An extended Kalman filter as in [7] is
used as well as nonlinear observer presented by [8] and [9].
The instantaneous cornering stiffness C () indicates when
y,max is reached, which is necessary for vehicle control.
As additional input y,max can be used in the -observer
as shown in Fig. 3. As discussed in [11], another potential
issue is the stability/convergence analysis in feeding back
y,max . Although not addressed in this paper, the problem
of the stability/convergence of the overall state estimation
(Fig. 3) has to be investigated in future works. For IGCC
the instantaneous cornering stiffness has to be estimated in
real-time. Therefore the algorithm has to work with limited
computational effort and only utilizing present and past data.
Two methods are presented for online estimation of C ():
Least Squares Online Estimation (III-C)
Weighted Recursive Least Squares Online Estimation
(III-D)
The offline estimation (III-B) is used as a reference to the
online methods.
B. Offline Estimation
For the offline estimation the whole data set can be used,
even the values of future vehicle states. Thus it is possible
to fit a tire model to the y --data set. Similar to the tire

4592

model presented in [10] the following more flexible structure

is used:
a2 + b
y = 2
(5)
c + d + 1

the instantaneous cornering stiffness C,i can now be written

as:

dy 
ref
d
(10)
C,i = C (i ) =
d =t

The four parameters of the tire model have to be optimized to

get a curve fitting regarding to total least squares (TLS) [12].
Because of different noise levels of y - and -data, a normalization has to be done before executing the algorithm.The
data has to be filtered with a high pass filter in order to get
the noise signal for each data. The standard deviation of
the noise signals is now used for the normalized coefficient
of friction ny :

(6)
ny = y
y

This method needs a lot of computational effort. On the one

hand, the fitting of the tire model to the measured data needs
many calculation steps. On the other hand, the calculation of
the related data points of the tire model has to be performed
for each data point. However, this method is reasonable as a
reference to benchmark the methods for online estimation.

normalized lateral coefficient of friction ny [-]

The n data points [P1d , P2d , ..., Pid , ..., Pjd , ..., Pnd ] relate

0.14

0.135

C. Online Estimation with LS

For this method a regression line is calculated for a defined
data window of the length l. The instantaneous cornering
stiffness C,i representing the i-th data point can be written
as slope of above mentioned regression line
Pi
(k
i )(y,k
y,i )
ls
(11)
C,i
= k=il
Pi
2
i)
k=il (k
with
i and
y,i being the mean value of the data set
[il , ..., i ] and [y,il , ..., y,i ]. It is necessary to use a
variable window length to have a satisfying performance at
any driving situation. The window length l is chosen minimal
fulfilling the condition

Pit .

0.13

0.125

max([il , ..., i ]) min([il , ..., i ]) min

Pid

0.12

0.115

Pjt

with min being constant. If min is chosen large, the

estimation will be slow especially at large values of d
dt .
Choosing a small value leads to a poor estimation being very
dependent on noise. For further consideration min should
be optimized by fulfilling an optimization criterion.

0.11

Pjd

0.105

D. Online Estimation with WRLS

0.1
0.04

0.045

0.05

0.055

0.06

0.065

0.07

0.075

0.08

tire slip angle [rad]

Fig. 4.

Projection of the data points orthogonally on the fitted tire model

to the n points [P1t , P2t , ..., Pit , ..., Pjt , ..., Pnt ] on the tire
model so that the line Pid Pit is orthogonal to the tire model
respectively to its slope at Pit as shown in Fig. 4. The
parameter vector is optimized iterative using a constant step
size. For all the permutations of the parameter vector the
criterion V of the summed squared orthogonal distances
V =

n
X

Pid Pit

(7)

i=1

In this section the regression line is calculated with the

WRLS-algorithm. The vector includes the two parameters
of the regression line and is calculated recursively with
xi = [i , 1] and yi = y,i , the forgetting factor , and the
covariance of the error P :
Pi1 xi
(13)
=
T
xi Pi1 +
1
Pi =
(1 xTi )Pi1
(14)

i+1 = i + (y xTi i )
(15)
wrls
C,i
is given by the first entry of i . Like in [13] a variable
forgetting factor is used. Similar to the window length l
in III-C the data points of the past are weighted by .
is assumed as proportional to l with a lower and upper
saturation:

has to be calculated. The parameter vector with minimum V

becomes the new vector for the next iteration step. So the
parameter vector converges to a vector fitting the tire model
to the data points. Considering the definition of
Pid
Pit

=
=

(12)

[di , dy,i ]
[ti , ty,i ]

and

(8)
(9)

= min
l lmin
= min +
lmax lmin
= max

for l lmin
for lmin < l < lmax (16)
for l lmax

The constraints max and min have to be chosen to have

an accurate estimation at l lmax and a fast performance at

4593

E. Estimation of y,max
As shown in Fig. 1 the maximum coefficient of friction
y,max is reached when C () = 0. Because of measurement noise this condition has to be relaxed in order to get a
robust y,max -detection (see Fig. 3):
C (i ) < C,crit

y,max,i = y,i

(17)

C (i ) C,crit

y,max,i = 1

(18)

instantaneous cornering stiffness C () [1/rad]

l lmin . Together with min used for the calculation of

l in (12) there are 5 free parameters for the optimization of
. These could be chosen by minimizing a criterion in order
to get optimal performance.

14

12

10

reference
LS
WRLS

4
0

10

12

14

16

time t [s]
ls ()- and C wrls ()-estimation method at a
Fig. 5. Validation of the C

circle movement on wet road

12

window length l [s]

The constant C,crit has to be chosen such that for given

vehicle data a robust detection results. As an input in an
-observer, y,max is only needed in the nonlinear region
of the tire model. In the linear region the cornering stiffness
C (0) is essential. Therefore y,max is assumed to be 1
according to (18), as long as data are only available for the
linear region of the tire model.

10

IV. EXPERIMENTAL RESULTS

The measurement was done with a modern passenger
vehicle having attached velocity sensors. Thus it is possible
to calculate with (4). In the first experiment the vehicle
was doing a circle movement on wet road. With increasing
speed and steering angle the vehicle starts to skid, reaching a
velocity of 20m/s. After reducing velocity and steering angle,
decreases abruptly reaching the monotone part of the tire
model.In Fig. 5 the reference Cref () starts at high level and
reaches the low level at about t = 6s, where C () becomes
zero or even negative. At t = 12s C () is changing rapidly
back to high level.
Both the LS- and WRLS-estimation do not reach the value
of the reference algorithm at high level, because of the effect
of the noise of . However, the high and low level can be
distinguished in their value. The LS-estimation follows the
reference-curve rapidly at changes, but it is varying strongly
at low level due to small window length l shown in Fig. 6.
The WRLS-algorithm causes a filtering, making noise of the
signal of Cwrls () very small.
Even though the forgetting factor is chosen variable as
shown in Fig. 7 the WRLS-estimation is not as fast as the
LS-estimation at the change of C () high to low level. It
is hard to chose the parameters of the variable forgetting
factor in (16) so that the estimator works well in every part.
That is the reason of errors in the constant high level area
at 0s < t < 6s and 13s < t < 15s. In the second experiment
the vehicle was absolving a slalom course on dry road at a
velocity of 20m/s. In Fig. 8 the LS- and WRLS-estimation
follow the reference only with a small time delay and the
low and high level areas can be separated. Because of the
fast change of in slalom maneuvers l and are very small.
The time delay is not caused by the computational effort of
the estimation, but because of the noisy data and the usage of

0
0

10

12

14

16

time t [s]
Fig. 6.

Window length l referring to the LS-estimation in Fig. 5

only the past data. However, the LS-estimation is sufficiently

fast and accurate for the use in vehicle control.
In Fig. 9 the validation of the y,max -estimation is shown.
The vehicle is moving along four different surfaces with
different values of y,max :
dry asphalt
wet asphalt
snow
ice

y,max
y,max
y,max
y,max

= 0.85
= 0.75
= 0.25
= 0.05

0s < t 12s
12s < t 24s
24s < t 36s
36s < t 60s

On dry and wet asphalt the vehicle is accelerating to 70km/h

with increasing steering angle. On snow the vehicle is doing a
slalom maneuver with 40km/h. On the icy surface the vehicle
is doing a cornering maneuver with 25km/h reaching a front
sideslip angle of 0.15rad. According to (18) y,max is set
to the value 1 when the vehicle is moving in the linear
region of the tire model, which is defined by C ()
C,crit , where C,crit = 1 was found to deliver good results.
The reference algorithm can clearly distinguish between the
surfaces and detect the correct value of y,max . The LSEstimation can not distinguish between dry and wet asphalt,

4594

but shows an accurate and fast behavior on snow and ice.

Values of y,max = 1 for short durations on wet asphalt
and snow are caused by short stretches of straight driving.
Under these driving conditions no information on the current
value of y,max is available and an estimation scheme for
the maximum longitudinal friction coefficient x,max would
have to be implemented. Because of the higher performance
of the LS-Estimation, the WRLS-Estimation is not shown
here.

forgetting factor [-]

1
0.995
0.99

0.985
0.98

0.975
0.97

0.965
0.96

0.955
0

10

12

14

16

time t [s]

instantaneous cornering stiffness C () [1/rad]

Fig. 7.

Forgetting factor referring to the WRLS-estimation in Fig. 5

The method described in III-B works well for the offline

estimation of the instantaneous cornering stiffness. The LSestimation is very flexible due to the variable window length
l and has a satisfying performance at both constant and
changing tire slip angle, using only one parameter for the
calculation of l. The forgetting factor of the WRLS-algorithm
can be tuned by 5 parameters, but the structure of the
variable forgetting factor is not flexible enough to satisfy
under all driving conditions. Therefore, the LS-estimation of
the instantaneous cornering stiffness was used for the y,max estimation and leads to accurate results.

30

reference

25

V. CONCLUSIONS AND FUTURE WORKS

A. Conclusions

LS
WRLS

20

15

B. Future Works

10

5
0

0.5

1.5

2.5

3.5

4.5

time t [s]
ls ()- and C wrls ()-estimation method at a
Fig. 8. Validation of the C

slalom movement on dry road

maximum friction coefficient y,max [-]

A suitable criterion can be incorporated into the computation of the variable window length l of the LS-estimation
and the optimal parameters of the forgetting factor of the
WRLS-estimation. Finding other structures of could lead to
a better performance as well. Since the output of the y,max estimation presented in this paper is still affected by noise,
an improvement can be expected by implementing a suitable
filter.
VI. ACKNOWLEDGMENTS
This research was part of the project SER 811111, which
is funded by FIT-IT, a program of FFG. The project is a
cooperation of Vienna University of Technology and TTTech
Computertechnik AG.

0.9

R EFERENCES

0.8

0.7

0.6

0.5

0.4

0.3

0.2

LS
reference

0.1

0
0

10

20

30

40

50

60

time t [s]
Fig. 9.

Validation of the y,max -Estimation on four different surfaces

4595

[1] F. Gustafsson, Slip-based Tire-Road Friction Estimation, Automatica, vol. 33 no. 6, 1997, pp 1087-1099.
[2] C. Lee, K. Hedrick, and K. Yi, Real-Time Slip-Based Estimation of
Maximum TireRoad Friction Coefficient, IEEE/ASME Transactions
on Mechatronics, vol. 9, no. 2, 2004
[3] T. Umeno, Estimation of Tire-Road Friction by Tire Rotational
Vibration Model, R&D Review of Toyota CRDL, vol. 37 No. 3, 2002
[4] W. Sienel, Estimation of the Tire Cornering Stiffness and its Application to Active Car Steering, Proceedings of the 36th Conference
on Decision and Control, 1997, pp 4744-4749
[5] R. Anderson and D. M. Bevly, Estimation of tire cornering stiffness
using GPS to improve model based estimation of vehicle states,
Proceedings of the IEEE Intelligent Vehicles Symposium, 2005, p 801806
[6] H. E. Tseng, Dynamic estimation of road bank angle, Vehicle System
Dynamics, vol. 36, no. 4-5, 2001, pp. 307-328
[7] L. R. Ray, Nonlinear State and Tire Force Estimation for Advanced
Vehicle Control, IEEE Transactions on Control Systems Technology,
vol. 3, no. 1, 1999, pp 117-124
[8] L. Imsland, T. A. Johansen, T. I. Fossen, J. C. Kalkkuhl, and A.
Suissa, Nonlinear observer for vehicle velocity estimation, SAE
International 2006

[9] A. von Vietinghoff, M. Hiemer, and U. Kiencke, Nonlinear observer

design for lateral vehicle dynamics, IFAC World Congress 2005
[10] U. Kiencke and A. Daiss, Estimation of Tyre Friction for Enhaced
ABS-Systems, Proceedings of the AVEG94, 1994
[11] M. Lakehal-ayat, H. E. Tseng, Y. Mao and J. Karidas, Disturbance
Observer for Lateral Velocity Estimation, Proceedings of AVEC, 8th
International Symposium on Advanced Vehicle Control, 2006
[12] S. Van Huffel and J. Vandewalle, The Total Least Squares Problems:
Computational Aspects and Analysis, Frontiers in Applied Mathematics, SIAM Philadelphia, 1991
[13] L. Ljung and T. Soderstrom, Theory and Practice of Recursive
Identification, MIT Press, Cambridge, Massachusetts, USA, 1983

4596