Vehicle System Dynamics

International Journal of Vehicle Mechanics and Mobility

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

Estimation of longitudinal speed robust to road

conditions for ground vehicles

Ehsan Hashemi, Alireza Kasaiezadeh, Saeid Khosravani, Amir Khajepour,

Nikolai Moshchuk & Shih-Ken Chen

To cite this article: Ehsan Hashemi, Alireza Kasaiezadeh, Saeid Khosravani, Amir
Khajepour, Nikolai Moshchuk & Shih-Ken Chen (2016): Estimation of longitudinal
speed robust to road conditions for ground vehicles, Vehicle System Dynamics, DOI:
10.1080/00423114.2016.1178391

To link to this article: http://dx.doi.org/10.1080/00423114.2016.1178391

Published online: 14 Jun 2016.

Submit your article to this journal

View related articles

View Crossmark data

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

http://www.tandfonline.com/action/journalInformation?journalCode=nvsd20

Download by: [University of Cambridge] Date: 15 June 2016, At: 00:23

 VEHICLE SYSTEM DYNAMICS, 2016
http://dx.doi.org/10.1080/00423114.2016.1178391

Estimation of longitudinal speed robust to road conditions for

ground vehicles
Ehsan Hashemia , Alireza Kasaiezadehb , Saeid Khosravania , Amir Khajepoura ,
Nikolai Moshchukb and Shih-Ken Chenb
a Department of Mechanical and Mechatronics Engineering, University of Waterloo, Waterloo, ON, Canada;
b General Motors Company, Global Research & Development, Warren, MI, USA

ABSTRACT ARTICLE HISTORY

Downloaded by [University of Cambridge] at 00:23 15 June 2016

This article seeks to develop a longitudinal vehicle velocity estimator Received 8 September 2015
robust to road conditions by employing a tyre model at each corner. Revised 5 April 2016
Combining the lumped LuGre tyre model and the vehicle kinemat- Accepted 7 April 2016
ics, the tyres internal deflection state is used to gain an accurate KEYWORDS
estimation. Conventional kinematic-based velocity estimators use Vehicle state estimation;
acceleration measurements, without correction with the tyre forces. uncertain dynamics; stability
However, this results in inaccurate velocity estimation because of analysis; parameter
sensor uncertainties which should be handled with another mea- estimation; unscented
surement such as tyre forces that depend on unknown road friction. Kalman filter
The new Kalman-based observer in this paper addresses this issue by
considering tyre nonlinearities with a minimum number of required
tyre parameters and the road condition as uncertainty. Longitudinal
forces obtained by the unscented Kalman filter on the wheel dynam-
ics is employed as an observation for the Kalman-based velocity
estimator at each corner. The stability of the proposed time-varying
estimator is investigated and its performance is examined experi-
mentally in several tests and on different road surface frictions. Road
experiments and simulation results show the accuracy and robust-
ness of the proposed approach in estimating longitudinal speed for
ground vehicles.

1. Introduction
Two major vehicle velocity estimation approaches are used in recent literature. The first one
uses vehicle’s planar kinematics, to develop observers [1,2] that use standard measurements
from an inertial measurement unit (IMU) such as acceleration, the yaw, the pitch, and the
roll rates, without implementing a tyre model. This approach is independent from tyre
parameters, but instead the sensors bias and noise should be identified precisely to have a
reliable estimation. Employing accelerations and wheel speed measurements, Imsland et al.
proposed a nonlinear observer in [3,4] to tackle model uncertainties. A Kalman observer
is employed by Hsu [5], utilising the same kinematic-based method and correction with
GPS data, but the poor accuracy of the mostly practiced conventional GPSs and the loss of
reception are primary concerns. The issue of reliance on a GPS is addressed with observer

CONTACT Ehsan Hashemi [email protected]

© 2016 Informa UK Limited, trading as Taylor & Francis Group

 2 E. HASHEMI ET AL.

development around tyre capacities and using the wheel speed, as investigated by Baffet
et al.,[6] but this approach is incapable of predicting the lateral limit before it occurs. This
drawback is eliminated in [7] by implementing a steer-by-wire system to determine the
road condition, tyre limits, and the slip angles of the front and rear tracks independently.
Yoon and Peng [8] utilise two low-cost GPS receivers for the lateral state estimation and
compensates the low update rate issue of the conventional GPS receivers by combining the
IMU and GPS data with an extended Kalman filter (EKF).
Another velocity estimation method uses a tyre model together with vehicle kinemat-
ics. However, this approach needs road condition information and tyre parameters, which
may vary significantly. The issue of time-varying tyre parameters is handled with a model
scheduling in [9], which again lacks information on road conditions. An EKF is employed
for both longitudinal and lateral state estimation in [10], but the load of computations
justifies using a reliable approach such as an unscented Kalman filter (UKF) without any
need for linearisation in the system dynamics. The UKF [11,12] and the simplified empir-
Downloaded by [University of Cambridge] at 00:23 15 June 2016

ical Magic formula [13,14] (as the tyre model) is employed in [15] for the velocity and
force estimation at each corner, but road friction and wheel torques are required for this
approach. Alternatively, a sliding-mode observer is proposed in [16], based on the LuGre
dynamic model,[17] to estimate the longitudinal speed, but the road should be identified
simultaneously. Zhang et al. presented a sliding-mode observer for velocity estimation in
[18], with the wheel torques and an EKF for estimation of the Burckhardt’s model [19]
parameters. However, this method needs an accurate tyre model, which is not possible due
to the presence of tyre wear, inflation pressure, and road friction uncertainties. Vehicle state
estimation on a single track car model is studied for slip angle estimation in [20], with a
Pacejka tyre model.[14] The derivatives of the lateral forces in their approach, however,
amplify noise effects in the lateral/longitudinal state estimates.
In this paper, kinematic equations are combined with the average lumped LuGre model
[21] to estimate the longitudinal velocity and the tyre internal deflection state at each corner
independently, using a Kalman filter (KF). Advantages of the developed velocity estimator
are its robustness to road conditions and insensitivity to reasonable changes in tyre param-
eters. The stability of the time-varying estimator and boundedness of the error covariances
are also addressed in this article. Furthermore, the stability of the estimator is studied for
deterministic and stochastic initial state vectors.
This paper provides different observers for the vehicle state estimation on various road
conditions with the following contributions: (1) longitudinal and vertical tyre-free force
estimators are developed using computationally efficient nonlinear observers and com-
mon measurements without knowing the road condition and any limiting assumption. (2)
A novel observer for the velocity estimation, robust to road friction changes is introduced
which treats acceleration measurement noises and the road condition as uncertainties.
(3) The corner-based structure of the longitudinal/lateral force and velocity estimators
advantageously leads to better performance of the stability and traction control systems
in conditions with different road friction under each tyre.
This article is divided into five sections. The longitudinal force estimation approach
is examined in Section 2, as are corner-based estimation by UKF and unknown input
observer (UIO) methodologies. Section 3 proposes a velocity estimator with a high-slip
detection module utilising the KF. Proof of the error covariance boundedness and the
stability of the linear time-varying estimator are also provided in Section 3. Road test
 VEHICLE SYSTEM DYNAMICS 3

experiments and simulation results are presented in Section 4, with discussion on the find-
ings in various manoeuvres with high and low longitudinal excitations. Section 5 includes
conclusions and future work.

2. Longitudinal force estimation

Tyre forces can be measured at each corner, but their cost impact, calibration and main-
tenance are their major drawbacks to be used for production vehicles. Provided that
the longitudinal tyre force calculation needs road friction, even accurate slip ratio infor-
mation from the GPS will not provide forces at each corner. Hence, estimation of tyre
forces independent of road conditions would be a remedy. Longitudinal force estimation
independent of the road friction may be classified on the basis of wheel dynamics into
the nonlinear and sliding-mode observers,[22,23,24] Kalman-based estimation,[25,26]
and UIOs.[27,28] This section provides two force estimation approaches, an UIO and a
Downloaded by [University of Cambridge] at 00:23 15 June 2016

Kalman-based method. The latter is selected in this study for the velocity estimation as it
is less sensitive to the tyre ’s effective radius and provides more smooth outcomes for the
transient regions during acceleration and brake. Experimental and simulation results are
presented in Section 4 to show the estimators performance in different manoeuvres and on
various road friction conditions.

2.1. Force estimation with UIO

A corner-based UIO method was proposed by Wang et al. ,[28] but the stability of the error
dynamics is explored in this paper. This method considers longitudinal tyre forces Fx as
unknown inputs for the wheel dynamics
1
ω̇ = (T t − Re Fx − Cb ω), (1)
Iw
where Re is the wheel effective radius, Tt represents the total effective torque on the wheel
at each corner, ω̇ is the wheel’s angular acceleration, and Cb shows the wheel bearing’s
linear viscous damping. The bearing viscous damping forces are negligible compared to
other terms in Equation (1). This approach provides a virtual wheel speed using a so-called
proportional, integral, derivative (PID)-type estimator ω̂˙ = (1/Iw )(Tt − Re f0 + Kp ω̃) in
which the constant f0 is the nominal value of the longitudinal tyre force and can be an
arbitrary bounded value. The difference between the measured and the virtually estimated
wheel speed (ω̃ = ω − ω̂) is used for the force estimation with the proportional term, as
in [28]
1
Fˆx (s) = f0 − Kp ω̃(s). (2)
Re
Matching between the measured (1) and virtually estimated wheel speed ω̂ means that
the control input produces the effect of unmeasured −Re Fˆx . Thus, the estimation error
dynamics for the wheel speed yields
ė = −Ke − B , (3)
where K = Kp /Iw and B = Re (Fx − f0 )/Iw . The stability of error dynamics (3) is investi-
gated in the following.
 4 E. HASHEMI ET AL.

Theorem 2.1: The UIO (2) for√ longitudinal force estimation is stable and the estima-
tion error is bounded by b/2K ξ(1 − ξ ) where 0 < ξ < 1 and b is the bound for B in
Equation (3).

Proof: To study the stability of the error dynamics (3), the Lyapunov function V = 12 e2 is
used. The derivation of the Lyapunov function yields

V̇ = −2KV − eB ≤ −2KV + e B . (4)

The term B is bounded because of the bounded characteristics of the real longitudinal
force. Therefore, B  < |b|, and the time derivative of the Lyapunov function is developed
again by introducing 0 < ξ < 1 as

V̇ ≤ −Ke2 + be
Downloaded by [University of Cambridge] at 00:23 15 June 2016

≤ −K(1 − ξ )e2 − Kξ e2 + be

 2
 b b2
≤ −2K(1 − ξ )V − Kξ e −  +
2 Kξ 4Kξ
b2
≤ −2K(1 − ξ )V + . (5)
4Kξ

The term 2K(1 − ξ ) is replaced with γ , and D stands for b2 /4Kξ ; then

V̇ + γ V ≤ D. (6)

Multiplying both sides of the inequality by eγ t and then integrating results in

 t
V − V(0) e−γ t ≤ D eγ (τ −t) dτ . (7)
0

Substituting the Lyapunov function V = 12 e2 in Equation (7) leads to

1 2 D
e ≤ V(0) e−γ t + (1 − e−γ t ). (8)
2 γ
Consequently, the bound for the error can be described as

e ≤ e(0)e−γ t/2 + 2D/γ . (9)
 √
The term 2D/γ is simplified to b/2K ξ(1 − ξ ), showing that any increase in K leads
to a decrease in e. As a result, the UIO for the longitudinal force estimator (2) with the
proportional term has a bounded estimation error and can be utilised whenever, wheel
torques are available. 

2.2. Kalman-based force estimation

Estimation problems can be addressed by UKF for the discrete-time nonlinear system of
the form xk+1 = F (xk , uk , f , npk ), yk = H(xk , f , nmk ). Parameters f and system states xk
 VEHICLE SYSTEM DYNAMICS 5

can be estimated recursively from the noisy output yk . Uncertainties in the process and
measurements are incorporated into the nonlinear system definition as npk , nmk . Proper
capturing of nonlinearities contributes to the unscented transformation that defines the
Sigma vectors X ∈ RN×2N+1 (N is the length of the state vector), which are supposed
to propagate through the nonlinear system. With some minor changes, UKF can also be
employed for parameter estimation instead of state estimation for the vehicle parameter
identification [29,30] and for the longitudinal force estimation.[26] For the force estimation
with UKF, the effective torque Tt provides input uk ; the wheel speed is assumed to be the
state xk , and the estimated longitudinal force Fˆx is denoted by the estimated parameter f̂ .
The discrete-time parameter estimation problem can be expressed as

fk+1 = fk + k , zk = G (xk , fk ) + νk , (10)

where zk corresponds to nonlinear observation on fk and k , νk represent process and

Downloaded by [University of Cambridge] at 00:23 15 June 2016

observation noises, respectively. In a parameter estimation problem, the estimated mean

is updated as f̂mk = f̂k−1 and initialised by f̂0 = E[f ]. The moving sample points, i.e.
sigma vector Fk | k−1 are generated around the estimated mean  f̂mk with the conven-
tional unscented transformation pattern Fk | k−1 = [f̂mk f̂mk + τ P̄fk f̂mk − τ P̄fk ], where
square root factorisation of the covariance matrix P̄fk is obtained by Cholesky decom-
position at each time step k. The error covariance matrix of the estimated parameter is
initialised with Pf0 = E[(f − f̂0 )(f − f̂0 )T ] and updated by P̄fk = Pfk−1 + Qk−1 with incor-
poration of the process noise covariance Qk−1 . Furthermore, τ is a scalar and represents
the spread of the sample points far from the mean values of random variables (states).
√
It is defined in [31] as τ = N + η, where η is the compound scaling parameter η =
2 N − N. Spread of the sample points around f̂ √
mk is denoted by = 3/N. Afterward,
β = 2 is introduced to employ the prior information on the Gaussian distribution of x.
Sample points are supposed to be propagated within the system (wheel dynamics) as
Zk | k−1 = G (xk , Fk | k−1 ), and the estimated function output ẑk is achievable from ẑk =
2N m
i=0 Wi Zi,k | k−1 . The weighting parameters are also defined by Wi = Wi = 2 (N + η)
c m 1

for all sets i ∈ {1, 2, . . . , 2N}. These parameters are W0c = η/(N + η) + 1 − 2 + β and
W0m = η/(N + η) for i = 0. The updated covariance matrices are given in Equation (11)
using the measurement noise covariance Rk


2N
Pzk zk = Wic (Zi,k | k−1 − ẑk )(Zi,k | k−1 − ẑk )T + Rk ,
i=0


2N
Pfk zk = Wic (Fi,k | k−1 − f̂mk )(Zi,k | k−1 − ẑk )T . (11)
i=0

The Kalman gain is achievable, by implementing these covariance matrices as Kk =

Pfk zk Pz−1
k zk
. As a result, the updated parameter and error covariance matrices can be obtained
as follows [31]:

Pfk = P̄fk − Kk Pzk zk KkT ,

f̂k = f̂mk + Kk (zk − ẑk ), (12)
 6 E. HASHEMI ET AL.

where f̂k is the updated longitudinal force estimate F̂xij at each corner. Outcomes of the UIO
and UKF approaches are compared in Section 4. The UKF moving sigma points through
the wheel dynamics greatly reduce the estimation fluctuations, even with the presence of
major uncertainties such as the road friction and changes in the effective radius. Therefore,
longitudinal force estimation with UKF is selected in this study and utilised for the velocity
estimation.
Estimated longitudinal forces should be normalised to be utilised in the velocity esti-
mator, which will be explored in the next section. In order to normalise the longitudinal
forces, normal (vertical) forces at each corner (tyre) should be calculated using lateral
and longitudinal vehicle dynamics. One can write the vertical and longitudinal accel-
eration components of the longitudinal dynamics as āθx = axm cos θv + azm sin θv and
āθz = azm cos θv − axm sin θv in which axm , azm are the measured longitudinal and vertical
accelerations by an IMU attached to the sprung mass. These measurements are affected by
Downloaded by [University of Cambridge] at 00:23 15 June 2016

the kinematics of the vehicle’s CG, vehicle pitch angle θv , and the road grade angle θr . The
sprung mass roll and pitch are not achievable by integration over the roll and pitch rate sig-
nals φ̇m , θ̇m because of sensor drift. Rehm provided a linear observer with low-pass filtering
in [32] to estimate the vehicle body’s roll/pitch angles as φ̂˙ = φ̇ + L e , θ̂˙ = θ̇ + L e
v m φ φ v m θ θ
by compensation over the error between the filtered estimates and the stationary roll/pitch
values, i.e. eφ , eθ and the observer gains Lφ , Lθ . Normal forces at front and rear axles thus
can be calculated by [24]

m m
F̂zf = − hCG āθ x + br āθz ,
Lwb Lwb
m m
F̂zr = hCG āθx + bf āθz , (13)
Lwb Lwb

where the subscripts f, r represent the front and rear axles, hCG is the height of the vehicle’s
centre of gravity, and the vehicle’s body pitch is denoted by θv . The wheel base is denoted
by Lwb = bf + br in which bf , br are the distances from the front and rear axles to CG.
Similarly, the vertical and lateral acceleration components of the lateral dynamics are āφy =
aym cos φv + azm sin φv and āφz = azm cos φv − aym sin φv in which aym is the measured
lateral acceleration by IMU which contains the kinematics of the vehicle’s CG, vehicle roll
φv angle, and the road bank angle φr . Therefore, using Equation (13) and equivalent masses
at each axle mf = Fzf /g, mr = Fzr /g, normal forces at each corner are as follows:

mi Ti
F̂ziL = āφz − hrc sin φ v − āφy hCG ,
Ti 2
mi Ti
F̂ziR = āφz + hrc sin φ v + āφy hCG , (14)
Ti 2

in which i ∈ {f, r} (front and rear axles), Tf , Tr represent the length of front and rear tracks,
respectively, and hrc is the height of the roll centre. Normalised longitudinal forces at each
tyre then can be written as μxij = F̂xij /F̂zij where j ∈ {L, R} shows the left and right tyres.
The normalised estimated forces μxij are used for the proposed velocity estimator and will
be described in the next section.
 VEHICLE SYSTEM DYNAMICS 7

3. Longitudinal velocity estimation

Kinematic-based velocity estimators and their limitations due to lack of observation are
investigated first in this section. Afterwards, a velocity estimation method that includes
observations on tyre forces is provided. However, the performance of the velocity estima-
tors based on tyre forces is practically limited because of uncertain road friction and tyre
parameters, especially for the transient and saturation regions. Consequently, this section
proposes a method that assumes road condition as uncertainty and estimates velocity with
a time-varying KF. The stability of such linear time-varying observer with deterministic
and stochastic initial conditions is also shown.

3.1. Kinematic-based velocity estimation

Because of difficulties in dealing with time-varying tyre parameters and unknown road
Downloaded by [University of Cambridge] at 00:23 15 June 2016

conditions, conventional kinematic-based velocity estimators employ acceleration mea-

surement and rely on GPS data intermittently. Linear, Kalman, or nonlinear observers
are used in such kinematic-based methods [1,5,6] without using a tyre model. Kinematic-
based estimation structure uses transformed kinematic equations at each corner with the
following dynamics:
u̇x = ax + ruy + axb + x,

u̇y = ay − rux + ayb + y, (15)

where the yaw rate is denoted by r, and the vehicle states ux , uy are estimated using the mea-
surements of acceleration and wheel speed. The bias and uncertainties in the acceleration
measurements are represented by axb , ayb and x , y , respectively. Any stochastic estima-
tor may be used for Equation (15) in which the only reliable and available measurement is
the wheel speed because of the lack of a tyre model which could provide tyre forces as an
additional measurement.
The integration of acceleration measurement, removing bias and noises by GPS, and
correction with the Kalman observer would consider providing vehicle states accurately,[5]
however, the accurate GPS data may not be available at all time and on production
vehicles due to high costs. Poor accuracy and low bandwidth of available commercial
GPSs, particularly in the lateral direction, and loss of reception are primary impediments.
Moreover, estimation errors for low-excitation conditions with continuous high-slip ratio
in such kinematic-based techniques makes the model-based estimation methods more
promising.[10,15,16] However, model-based approaches based on the correction with the
tyre forces should deal with the varying model (tyre) parameters and unknown road fric-
tion. This issue is tackled in the next section using the lumped LuGre tyre model without
the road friction term.

3.2. Corner-based velocity estimation

Implementing a tyre model independent of road conditions is desirable for velocity esti-
mation. Doing so is possible with use of an observer-based approach, which treats the road
friction as a bounded uncertainty. This method is targeted in this article and experimen-
tally tested for both dry and slippery road conditions. The average lumped LuGre model is
 8 E. HASHEMI ET AL.

introduced in [21] to obtain the tyre forces using the force distribution coefficient κ, rather
than by integrating bristle element forces over the tyre patch length. Representing average
deflection of the bristles, the tyre internal state z(ζ , t) in the average lumped LuGre model,
has the capability of being used in a velocity estimator as a state, because of its dynamics

σ0l |Vrl |
żl (ζ , t) = Vrl − κl Re |ω| + zl (ζ , t), (16)
θ g(Vrl )

in which l ∈ {x, y}, ζ is the axis coordinate, σ0l is the rubber stiffness, Vrl shows the relative
velocities. the level of tyre and road adhesion is represented by introducing a so-called
road classification factor, which may vary between 0 < θ ≤ 1 according to dry, wet, and icy
conditions, respectively. The longitudinal relative velocity can be replaced by Vrx = Re ω −
uxt where the longitudinal velocity in the tyre coordinate system is denoted by uxt . The
Downloaded by [University of Cambridge] at 00:23 15 June 2016

longitudinal relative velocity in the tyre coordinate of the LuGre model resembles the slip
ratio λ = (Re ω − uxt )/max{Re ω, uxt } in the mostly used tyre models such as Burckhardt
[19] and Pacejka and Besselink.[33] The parameter κ in the average lumped model can be
a constant, or a function of time, or may be approximated by an asymmetric trapezoidal
scheme. The suggested value for κ in [34] is κ = 76 L, where L is the tyre patch length.
α
In addition, g(Vrl ) is defined by g(Vrl ) = μc + (μs − μc ) e−|Vrl /Vs | , in which μc , μs are
the normalised Coulomb friction and static friction, respectively. The Stribeck velocity
Vs shows the transition between these two friction states. The tyre parameter α = 0.5 is
assumed for this study.
The final form of the normalised friction force (μl = Fl /Fn ) of the average lumped
LuGre model yields μl = σ0l zl + σ1l żl + σ2l Vrl . In addition, the internal state zl can be
expressed using the two-dimensional combined-slip LuGre model, which considers the
effect of longitudinal/lateral slip on the tyre capacity in another direction. The LuGre non-
measurable longitudinal internal state can be written at each corner (wheel) in the presence
of uncertainty z (t) as

żx (ζ , t) = Vrx − κx Re |ω|zx (ζ , t) + z (t). (17)

Uncertainty z (t) is replaced with the road friction term (σ0x |Vrx |/θ g(Vrx ))zx and is
bounded because of the passivity of the LuGre model. Moreover, the derivative of the
relative velocity is also corrupted due to the sensor noise and bias [35]

V̇rx = Re ω̇s − u̇xt + a (t), (18)

in which ω̇s stands for the wheel’s rotational acceleration and u̇xt represents the projected
longitudinal acceleration in the tyre coordinate system. The term a (t) shows the devia-
tion of the measured relative acceleration Re ω̇s − u̇x t from V̇rx because of the sensor noises.
Establishing these equations allow the development of a Kalman-based observer to incor-
porate longitudinal tyre deflections (17) and relative velocities (18) concurrently and to
compare the output with the longitudinal forces μ̂x provided in Section 2. The general
form of the velocity estimator at each corner ij, which can be addressed by the KF, with
 VEHICLE SYSTEM DYNAMICS 9

ω > 0 is given as follows in which  = [ z

T
a] :

ẋ = A(t)x + B(t)u +
⎡ ⎤⎡ ⎤ ⎡ ⎤
−κx Re ω 1 0 zx 0
= M −1 ⎣ 0 0 0 ⎦ ⎣Vrx ⎦ + M −1 ⎣1⎦ (Re ω̇s − u̇xt ) + ,
−κx Re ω̇ 0 −κx Re ω żx 0
y = Cx +  = σ0x zx + σ2x Vrx + σ1x żx + . (19)
Process and measurement noises are denoted by , , respectively, and M = [1 0 0; 0 1 0;
0 −1 1]. The linear time-varying system (19) uses a reduced number of tyre parame-
ters: normal force distribution factor κ, rubber stiffness σ0x , rubber damping σ1x , and
relative viscous damping σ2x . These tyre parameters are not related to the road con-
dition and friction parameters. The bounded error covariance and stability of linear
Downloaded by [University of Cambridge] at 00:23 15 June 2016

time-varying estimators for both known zero and nonzero initial error covariance were
explored in [36,37].
Observability is a sufficient condition for implementation of an optimal variance filter
(such as a Kalman estimator). On the other hand, the controllability property guarantees
asymptotic stability of such filters for the linear time-invariant case. The observability and
controllability conditions are next studied for the Linear Time-Varying (LTV) system (19),
which has the discrete-time form
xk+1 = Ak xk + Bk uk + k ,
yk = Ck xk + νk , (20)
with process and measurement uncertainties k , νk , which have the covariance Q̄k =
E[k , k T ], R̄k = E[νk , νk T ] accordingly. Process and measurement noises are assumed to
be uncorrelated E[k , νkT ] = 0 and have zero mean E[k ] = E[νk ] = 0; ∀k ∈ N.
Discretisation of the system is done by the step-invariance method, because of its pre-
cision and response characteristics, which prohibit divergence of the integration of stable
dynamic systems.[38] Input to the continuous-time system is the hold signal uk = u(tk ) for
a period between tk ≤ t < tk+1 with the sample time Ts . Then, the discrete-time system is
defined by xk+1 = Ak xk + Bk uk with state transition and input matrices Ak = eA(t)Ts and
T
Bk = 0 s eA(t)τ B(t) dτ .
The discrete-time Kalman observer suggests the following prediction with correction to
estimate the states (tyre internal state z and the longitudinal relative velocity Vrx ) defined
by x̂k+1 | j  E[xk | yj ] using a sequence of measurements yj
x̂k+1 | k = Ak x̂k | k−1 + Bk uk + Kk (yk − Ck x̂k | k−1 ), (21)
where the optimal Kalman gain is Kk = Ak P̄k | k−1 CkT (Ck P̄k | k−1 CkT + R̄k )−1 and error
covariance P̄k+1 | k  cov(xk+1 − x̂k+1 | k ) forms a discrete time-varying Riccati equation (22)
for both zero and nonzero state initialisation x̂0 | −1 = E[x0 ] and covariance initialisation
P̄0 | −1  cov(x0 ) = E[(x0 − x̂0 | −1 )(x0 − x̂0 | −1 )T ]
P̄k+1 | k = Ak P̄k | k−1 ATk + Q̄k − Kk Ck P̄k | k−1 ATk . (22)
The Kalman gain and error covariance do not depend on the measurements, even for
the time-varying case, but only on the noise statistics. The estimation error is defined by
 10 E. HASHEMI ET AL.

ek+1 | j  xk+1 − x̂k+1 | j , which yields

ek+1 | k = (Ak − Kk Ck )ek | k−1 − Kk νk + k . (23)

Stability of the estimator is investigated in the following.

3.3. Stability of the estimator

Stability of the error dynamics (23) is studied in this section with known and uncertain
initial conditions. The detectability and stabilisability definitions in Equations (30) and (31)
are required for the stability analysis of the suggested discrete-time estimators. Uniform
detectability leads to bounded error covariance. In addition, stabilisability of the paired
state transition matrix and process noise results in exponential stability of the estimator, as
proved in [39,40].
Downloaded by [University of Cambridge] at 00:23 15 June 2016

The stochastic observability, stability and convergence of the state mean, and bounds
on error covariance of the Kalman estimator for LTV systems, such as that in Equation
(21), were studied in [39,41]. These studies were focused on systems with determinis-
tic parameters and known initial state vectors and done in terms of uniform complete
observability and controllability grammians. On the other hand, the bounded error covari-
ance and stability of the Kalman estimator for systems with completely uncertain initial
covariance/states is investigated in [42].
Uniform detectability and stabilisability conditions are investigated in this section to
check the stability and error covariance boundedness of the proposed velocity estimator
for two cases: (a) known zero/nonzero initial states and (b) complete uncertainty of the
initial state statistics.

Proposition 3.1: There exists a state estimator such as Kalman (21) having bounded error
covariance for time-variant system (19) with deterministic time-varying parameters and
known initial state/covariance.

Proof: For the system xk+1 = Ak xk + k , yk = Ck xk + νk with uniform detectability of

[Ak , Ck ], the known initial state/covariance, and the process and measurement noise
covariances Q̄k , R̄k , there exists a state estimator such as the Kalman having bounded error
covariance (see appendix).[39] Furthermore, stabilisability of the pair [Ak , Gk ] leads to
exponential stability of the KF, where Gk is an appropriate matrix obtained by Qk = Gk GTk .
Proof is provided in [39].
Therefore, the detectability condition (30) should be examined for the proposed veloc-
ity estimator (19). This is experimentally checked for the deterministic time-varying wheel
speed of the discrete-time systems matrix Ak . The rank of V(0, N) on several road exper-
iments is shown in Figure 1, where N represents the total number of iterations with the
sample time Ts = 0.005 s.
Figure 1 shows that the suggested estimator has full rank on the observability gram-
mian (30) for all performed manoeuvres, showing the observability of the system for
known initial covariances. Manoeuvres include double lane change (DLC); acceleration-
in-turn (AiT); brake-in-turn (BiT); and launch on dry, snowy, and icy roads. The launch
manoeuvres do not have the full rank condition for t ≤ 2 because the wheel speed has
 VEHICLE SYSTEM DYNAMICS 11

Figure 1. Observability test: rank of V(0, N) for known initial state/covariance.

Downloaded by [University of Cambridge] at 00:23 15 June 2016

been zero for that period, which is not the concern of this study. The stabilisability con-
dition is also required since states should be affected by the noise such that the optimal
Kalman estimator is forced to utilise measurements. This condition is also satisfied and the
grammian (31) has full rank. Additionally, since the system matrix A(t) in Equation (19) is
physically bounded (because of the wheel speed and its derivative characteristics), a con-
ventional observability test is performed here. The observability matrix for system (19) is
given by [43]

On = [τ1 τ2 ··· τn ]T ,
τ1 = C, τi+1 = τi A(ω) + τ̇i . (24)

Observability is confirmed by holding the full rank condition rank(O3 ) = 3 at each fixed
time span for operating regions of the wheel speed and its time derivatives. Thus, the sug-
gested parameter-varying corner-based estimator (19) is observable and it is feasible to
observe the longitudinal tyre internal states ẑx and the relative velocity V̂rx with any known
initial covariance matrix and by employing the longitudinal force as the output. 

In the case of complete uncertainty on the initial state/covariance, the estimated covari-
ance matrices can be unbounded even if the LTV system satisfies the observability crite-
ria (30). The bounded error covariance and stability of the KF for the proposed velocity
estimator (19) with completely uncertain initial state/covariance is investigated here in the
following.

Proposition 3.2: The longitudinal states zx , Vrx of the time-varying system (19) can be
recovered using measurements μxij and a Kalman estimator with stochastic initial covari-
ance/states at each corner.

Proof: By definition, the system (20) is stochastically observable if there exists a finite time
tf , such that the state covariance matrix P̄k is bounded [42]

λmax (P̄k ) < λb , tk ≥ tf , (25)

where λmax (P̄k ) shows the largest singular value of the matrix P̄k and λb is a predefined
scalar bound. Assuming initial state covariance matrix P̄0 | −1 = ψI, ψ ∈ R, ψ > 0, one
 12 E. HASHEMI ET AL.

can rewrite the time-varying Riccati equation (22) as [42]

P̄k+1 | k = ψ Mk+1 + Nk+1 + Sk+1 , (26)

where Nk+1 = Nk+1 (Mk , Nk , φk , Ck Q̄k , R̄k ), Sk+1 = Sk+1 (ψ, Sk , Nk , φk , Mk , Ck , Q̄k , R̄k ),
and Mk+1  φk,0 X0,k X0,kT φ T . The procedure for obtaining X
k,0 0,k is provided in the
appendix and φi,j = φi,i−1 φi−1,j are the state transition matrices for i ≥ j with φi+1,i = Ai .
In summary, the following Lemma presents two tests for observability of the velocity
estimator with stochastic initial conditions.

Lemma 3.3: The Kalman estimator (21) on the system (20) with an error covariance matrix
(22) and stochastic initial state P̄0 | −1 = ψI, ψ ∈ R+ is stochastically observable if the condi-
tion λmax (Mf ) = 0 (test 1) holds for a finite time tf and λmax (Nk+1 ) < λb for tk ≥ tf (test 2)
with a predefined bound λb , where Mk+1 is obtained from the modified Riccati equation (26)
Downloaded by [University of Cambridge] at 00:23 15 June 2016

and the procedure provided in Equations (35) and (36) in the appendix. Employing the con-
dition λmax (Mf ) = 0 for a finite time tf < ∞, the modified Riccati equation (26) changes
to P̄k+1 | k = Nk+1 + Sk+1 which leads to a simplified form of Nk+1 as in

Nk+1 = Ak Nk ATk + Q̄k − Ak Nk CkT −1 −T

k k Ck Nk Ak ,
T
(27)

where k is expressed in Equation (36).

Whenever the two criteria on Mf , Nk+1 in Lemma 3.3 (so-called test 1 and test 2) are
met, the Kalman observer (21) is stable even if the scalar ψ has infinite values. These two
tests have been performed on the proposed observer with Q̄ = 2.5e − 3 and R̄ = 1e − 3
and results are depicted as follows. Figure 2 exhibits λmax (Mf ), where Mf is obtained
from Equation (26) and the procedure provided in the appendix. Different experiments
such as DLC, BiT, sharp turn, and steering on dry and slippery (snow/ice) roads have been
performed and results are illustrated in Figure 2.
From the plots in Figure 2, it is apparent that the largest singular value of Mf converges
to zero after tf = 0.03s for different experiments. The values of λmax (Nk+1 ) with Nk+1
from Equation (27) are plotted in Figure 3.
The results on different road conditions and pure/combined-slip manoeuvres, as shown
in Figure 3, indicate that the maximum singular value of Nk+1 remains bounded. Thus,

Figure 2. Test 1 for various road tests, experimental results.

 VEHICLE SYSTEM DYNAMICS 13

Figure 3. Test 2 for different driving scenarios and roads, experimental results.
Downloaded by [University of Cambridge] at 00:23 15 June 2016

both the criteria ( test 1 and test 2) are met and the discretised form of system (19) with KF
estimator (21) is stochastically observable. 

Consequently, the presented model-based estimation is stable, and errors of the state
mean have bounded variance for both known and stochastic initial covariance.
Figure 4 shows the structure of the discussed estimator, in which subscript ij represents
the estimation/measurement at each corner. The newly proposed longitudinal state esti-
mation approach, which is independent from the road condition, is developed using the
UKF force estimator (in Section 2) and the Kalman observer on the linear time-varying
system (19).
The presented UKF force estimator structure in Figure 4 depicts the procedure for the
sigma points propagation and the covariance matrix update. The estimated relative lon-
gitudinal velocities V̂rxij at each corner from Equation (21) are used for the longitudinal
velocity estimation at the tyre coordinates as ûxtij = Re ωij − V̂rxij . Afterward, each corner’s
longitudinal velocity in the vehicle coordinates ûxij yields

ûxij = ûxtij cos δ − ûytij sin δij , (28)

in which δij is the steering angle at corners and the estimated lateral velocity at each
corner’s tyre coordinates is denoted by ûytij . Consequently, longitudinal velocity at the
vehicle’s centre of gravity can be expressed as ûx = η1 ûxf + η2 ûxr in terms of corners’ esti-
mated velocities in which the average estimated velocities on the front axle are defined by
ûxf = (ûxfL + ûxfR )/2. Front and rear axle gains are denoted by η1 , η2 . The allocated gains
for the road experiments in this study are η1 = η2 = 0.5, but they can be variable and
defined adaptively based on the level of excitation.
A high-slip detection algorithm is also used to deal with the noises associated with
large slip ratio conditions. Noise covariance matrices change appropriately upon detec-
tion of a high-slip case to incorporate changing in the level of reliance on the vehicle
kinematics (process) and longitudinal forces (measurement). Covariance matrices Q̄, R̄
change adaptively to avoid fluctuations (caused by nonlinearities/disturbances) during
harsh manoeuvres on slippery surfaces.
This algorithm needs a slip ratio threshold λ̄th after which the process and measurement
covariance matrices change to Q̄ = 3.3 e−4 and R̄ = 2.1 e−1 , respectively. Sudden changes
 14 E. HASHEMI ET AL.
Downloaded by [University of Cambridge] at 00:23 15 June 2016

Figure 4. Structure of the longitudinal velocity estimation with the tyre model.

in the slip ratio (vehicles response) will not be detected in case of large constant high-
slip threshold. This leads to more required time for the estimated slip ratio to satisfy the
threshold (i.e. it requires larger excitations). On the other hand, small constant threshold
results in unnecessary detection of the large slip cases. Thus, in the developed high-slip
detection module, the threshold changes between the predefined upper and lower bounds
λu , λl according to the driving conditions as

1
√ e−βe σe /2ϕe ,
2
λ̄th = λu − (29)
ϕe 2π
√
where ϕe = (1/ 2π)(λu − λl ) and σe represents variance of the vehicle’s acceleration
ak over a moving window with  size Na , i.e. σe = var{ak 2 : m − Na ≤ k ≤ m}, ∀m ∈
N, m ≥ Na , in which ak 2 = a2xk + a2yk and axk , ayk are measured longitudinal and
lateral accelerations. The rate of transition between the predefined upper and lower thresh-
olds λu , λl is denoted by βe . Thus, for |λ| ≥ λ̄th the covariance matrices Q̄, R̄ change to the
new values.
To use the current measurement and remove the outliers and detect the large slip sce-
narios, time derivative of the wheel speed |ω̇| is also implemented and a threshold |ω̇th| is
introduced as a criterion after which the changes in the covariance matrices should happen.
 VEHICLE SYSTEM DYNAMICS 15

Finally, the covariance matrices change when each of the slip-based or wheel speed-based
criteria are met, i.e. {|λ| ≥ λ̄th ∨ |ω̇| ≥ |ω̇th|}. This leads to the prompt detection and con-
sequently proper covariance matrix allocation. Simulation and experimental results in the
next section confirm the validity of the algorithm on dry and slippery roads.

4. Results and discussion

This section includes simulation and experimental tests for validation of the longitudinal
state estimator on a fully electrified sport utility vehicle (SUV) with four independent wheel
drive, specifications given in Table 1, and longitudinal tyre parameters σ0x = 612, σ1x =
0.87, σ2x = 0.0016, κ = 8, and Re = 0.35.
For the purpose of analysis, road tests are performed, and results of the longitudinal
velocity and force estimators are presented for different road frictions. Measured signals
Downloaded by [University of Cambridge] at 00:23 15 June 2016

are communicated using a CAN-bus, a conventional communication method in produc-

tion vehicles. Real-time logging of this sensory information was done using the dSPACE
MicroAutobox . The dSPACE compiles measurements for MATLAB/SIMULINK, and
the controller provides control signals for the dSPACE as well. The test platform SUV
and input/output layout are shown in Figure 5. To validate the estimated forces, this vehi-
cle is equipped with additional sensors for direct measurement of tyre forces and moments
at each corner.
Several driving scenarios such as acceleration/deceleration, lane change (LC), DLC, AiT,
brake-in-turn (BiT), harsh steering, and launches were examined on various roads.

4.1. Force estimation results

In Figure 6(a) forces estimated by both the UKF and the UIO are compared with the Car-
Sim tyre forces on a slippery road for a manoeuvre with successive accelerator and brake
pedal requests. The simulation occurs on a slippery road with μ = 0.3, and the initial
longitudinal velocity of the vehicle is u0x = 30 kph.
The proposed force estimation methods are independent of the road condition and
the UKF estimator exhibit more accurate and smooth performance than UIO for such
manoeuvre with successive sign changes of the slip ratio. An AiT manoeuvre on dry
and slippery roads is simulated in the CarSim and results are graphically illustrated in
Figure 6(b).
For the AiT driving scenario, the accelerator is applied to 100% at t = 2 s and continues
till t = 4 s. It is released between 4 and 6 s and pushed up to 100% again as a step signal

Table 1. Vehicle specifications for simulation and experiments.

m (kg) 2270 Vehicle mass
Iz (kg m2 ) 4650 Vehicle moment of inertia
Iw (kg m2 ) 1.68 Wheel moment of inertia
Lwb (m) 2.85 Vehicle wheel base
bf , br (m) 1.42, 1.43 Front and rear axles to CG
Re (m) 0.33 Effective radius
hrc (m) 0.54 Roll axis height
hCG (m) 0.65 CG height
T f , Tr (m) 1.62, 1.56 Front and rear track width
 16 E. HASHEMI ET AL.
Downloaded by [University of Cambridge] at 00:23 15 June 2016

Figure 5. The fully electrified 4WD test platform and I/O layout.

till t = 11 s, then it is linearly reduced to zero at t = 15 s. A steering angle δsw = 1 rad is

imposed between t = 2 − 12 s as well and the steering ratio is rδ = 16.7. The performed
combined-slip AiT manoeuvre is harsh, but the suggested UKF for estimator by weighted
averaging of the sigma points’ values handles the oscillations in the transient regions
resulted from imposing and releasing torques on wheels better than UIO. The fluctuations
observed in the CarSim’s force profile curves are attributed to the requested acceleration
with high magnitude.
The road experiment results of the force estimation at the rear-left corner in a driving
condition with harsh steering on a wet road (μ ≈ 0.45) are shown in Figure 7 and com-
pared with the measurement for the test platform vehicle with the stability control system,
which prevents high side slip angles.
The effective torque and the wheel speed of the rear wheels Ttrj , ωrj are also depicted
in Figure 7. UKF-based and UIO estimators suggest accurate estimation for such harsh
steering condition. Another test, a DLC manoeuvre, with high-slip and lateral excitation is
conducted on a snowy surface and force estimation results of UIO and the proposed UKF
approach are compared in Figure 8 for the rear-left wheel. The force peak values in the
DLC manoeuvre (Figure 8) on snow is higher than the wet surface (Figure 7), which shows
that the driving on the wet road was not very severe.
There is a certain level of correspondence between the outcomes of the two estima-
tion methods and the measurements, even with presence of high slips. However, but UKF
exhibits more consistent and smooth results for the DLC on snow from t = 4.1 to t = 4.5 s
and around t = 5.1, 7.8 s which are the sudden drop or increase in the torques at each
corner.
 VEHICLE SYSTEM DYNAMICS 17

(a)

(b)
Downloaded by [University of Cambridge] at 00:23 15 June 2016

(c)

Figure 6. Estimated longitudinal forces in (a) acceleration/brake on a road with μ = 0.3, (b) AiT on dry,
and (c) AiT on a slippery road with μ = 0.25.

Accuracy of the longitudinal force estimators are evaluated in different manoeu-

 with the normalised root mean square of the error NRMS defined by ς̄ =
vres
Np
i=1 (p̂i − pi ) /Np /pm where the measured and estimated signals are denoted by p and
2

p̂, respectively, Np is the number of collected signal samples during a driving scenario
(DLC, AiT, BiT, etc.), and pm = max |pi | shows the maximum value of the measured
i=1···Np
signal. The NRMS ς̄ of the estimated longitudinal and vertical forces and velocities in dif-
ferent driving scenarios and on various road frictions are listed in Table 2 where ς̄1 and ς̄2
represent the NRMS for the UKF and UIO, respectively.
Table 2 substantiates that the NRMS of the estimated longitudinal forces by UKF is better
than the UIO and it is less than 6.25% for the performed manoeuvres on dry, wet, and
snowy roads. Therefore, the UKF approach is selected as the longitudinal force estimator
for the velocity estimation in Section 3 because of its superior performance. The moving
sigma points (in the UKF approach) through the wheel dynamics (1) reduce estimation
fluctuations, especially during transient regions, even in the presence of uncertainties such
as in road conditions, which may vary from icy to dry (i.e. 0.1 ≤ θ ≤ 0.97) in the tyre
 18 E. HASHEMI ET AL.

(a)

(b) (c) (d)

Downloaded by [University of Cambridge] at 00:23 15 June 2016

Figure 7. Road experiments, harsh steering, and LC on wet (a) estimated F̂x at rL, (b) rear wheel torques,
(c) rear wheel speeds, and (d) steering wheel angle δsw .

(a)

(b) (c) (d)

Figure 8. DLC on snow and experimental results (a) estimated F̂x at rL, (b) rear wheel torques, (c) rear
wheel speeds, and (d) steering wheel angle δsw .

model, effective radii with ±5% variation, and corrupted measurements of wheel speed
and torque with variance Rω = 0.18 and RTt = 32, respectively.
The vertical force estimator leads to the normalised error ε ≤ 3.9% which confirms
effectiveness of the algorithm on dry and slippery roads. Observed errors between the
measured and estimated forces in Table 2 may have several sources such as camber angle,
which has not been modelled in the estimation algorithm. Moreover, inaccurate inertial
parameters and uncertainties in the CG location contribute to such errors.
 VEHICLE SYSTEM DYNAMICS 19

Table 2. NRMS of the errors for the longitudinal force estimator (with ς̄1 , ς̄2 for the UKF and UIO) and
vertical force estimator.
Steer-wet/dry DLC-snow BiT/Accel-snow
Estimated forces ς̄1 (%) ς̄2 (%) ς̄1 (%) ς̄2 (%) ς̄1 (%) ς̄2 (%)
FxfL 3.85 4.05 4.87 6.11 5.04 4.90
FxfR 4.02 3.91 5.32 5.18 6.23 6.28
FxrL 3.97 4.26 4.76 5.22 5.77 6.33
FxrR 4.51 4.71 4.95 5.49 5.64 7.10
ς̄ (%) pm (kN) ς̄(%) pm (kN) ς̄ (%) pm (kN)
FzfL 3.34 12.77 3.65 9.91 3.88 9.10
FzfR 2.72 12.41 3.01 10.16 3.03 7.06
FzrL 2.15 10.02 2.74 8.37 3.54 7.19
FzrR 1.93 10.14 2.82 8.39 2.92 6.82
Downloaded by [University of Cambridge] at 00:23 15 June 2016

Figure 9. Estimated longitudinal velocity for SS and AiT on dry and slippery roads.

4.2. Velocity estimation results

The stochastic observability of the suggested velocity estimator (19) was studied in
the previous section for both known and uncertain initial states/covariances, and it is
experimentally concluded that for the case with uncertain initial covariance, in addition
to the observability condition, two supplementary criteria λmax (Mf ) = 0 (test 1) and
λmax (Nk+1 ) < λb (test 2) are met for various manoeuvres and road frictions.
To verify the proposed velocity estimator in manoeuvres with lateral excitation, a step
steer (SS) scenario is simulated in CarSim with the initial speed ux0 = 60 kph. Figure 9
illustrates longitudinal velocity estimates for this SS case on a dry road with steering wheel
angle δsw = 2 rad at t = 2 s. The simulation confirms that the newly proposed corner-
based velocity estimator provides accurate results in manoeuvres with both longitudinal
and lateral slips on dry road conditions.
The second set of analysis on the velocity estimator examines the suggested approach
in CarSim for an AiT manoeuvre. Figure 9 also demonstrates an AiT test with the initial
velocity of ux0 = 30 kph and maximum steering wheel angle δsw = 1 rad at t = 2 s on dry
and slippery (μ = 0.3) roads. Accelerator is applied to 100% at t = 5 s and continues till
t = 10 s.
 20 E. HASHEMI ET AL.

(a)

(b) (c)
Downloaded by [University of Cambridge] at 00:23 15 June 2016

Figure 10. Launch and AiT on wet and transition to dry (a) estimated speed, (b) measured accelerations,
and (c) measured yaw rate.

Simulation results reveal that the proposed estimator performs well on various road
conditions for manoeuvres with both longitudinal and lateral excitations. Outcomes of
the road experiments on a fully electrified SUV (shown in Figure 5) are presented in
the followings. One of the main objectives of such an estimator is to provide reliable
longitudinal velocity ûxt at each corner for traction control systems during launch (or
hard acceleration) on slippery roads. Figure 10(a) shows performance of the proposed
estimator for a launch on a highly slippery wet surface with μ ≈ 0.3 which ended on
a dry surface with a break. Moreover, an AiT manoeuvre on the same wet surface with
a break on dry was performed and estimated longitudinal speed results are provided in
Figure 10(a).
The input torque from the driver for the launch manoeuvre on highly slippery wet
surface brings the tyre up to their longitudinal capacity. The measured longitudinal acceler-
ation shown in Figure 10(b) for this manoeuvre is below 3.3 m/s2 which confirms slippery
conditions according to the required accelerator pedal by the driver up to the tyres ’ lim-
its. For both tests, the stability and traction controllers were active, but intentionally set to
have a poor performance, which leads to sudden increase in the wheel speed and subse-
quently slip ratio at each corner. Front tyre loose grip in launch and acceleration cases due
to the drop in the vertical force on the front track by the load transfer. Therefore, in such
manoeuvres high-slip ratio for the front tyres is a concern for traction control systems.
Wheel speed for the front tyres, i.e. fL, fR are shown in Figure 11 and compared with the
estimated velocities and the measured velocities by GPS.
 VEHICLE SYSTEM DYNAMICS 21

(a)

(b)
Downloaded by [University of Cambridge] at 00:23 15 June 2016

Figure 11. Wheel speed, measured corner velocities, and estimated velocities at wheel centres for (a)
launch on wet then dry and (b) AiT on wet then dry.

Results of such manoeuvres corroborates that even with the presence of high slips, the
proposed estimator provides accurate and reliable longitudinal velocity estimates ûxtij at
each tyre (wheel centre) and subsequently at the vehicle CG, i.e. ûx .
In order to assess the proposed approach in road experiments with high lateral exci-
tation (combined-slip characteristics) and large longitudinal slip, a BiT accompanied by
hard acceleration on snow (with μ ≈ 0.35) is done and the estimation results are provided
in Figure 12. Measured accelerations and the yaw rate for such BiT case is provided in
Figure 12(b) that shows the weak grip condition for each tyre.
Another experiment with harsh steering on snow and ice road (with μ ≈ 0.25) was exe-
cuted to validate the method and the outcomes are also demonstrated in Figure 12 together
with the measured accelerations and the yaw rate. As can be seen from Figure 12(a),
the developed algorithm with the high-slip detection module provides accurate velocity
estimates in manoeuvres with combined-slip characteristics on snowy and icy roads as well.
Wheel speed of the front tyres’ centre for the BiT and steering manoeuvres on snow/ice
are compared with the estimated and measured velocities in Figure 13. The estimates by
the proposed corner-based approach have correspondence with the measurement in spite
of the large slip cases around t = 6 s and after t = 11 s for the harsh steer on snow/ice and
in 4 ≤ t ≤ 7.8 s for BiT on snow.
The high-slip detection module changes the covariance matrices of the Kalman-based
estimator based on the methodology discussed in Section 3. This significantly improves
the outcomes by defining the level of reliance on the forces for the velocity estimation
correction.
The simulation and experimental results provided in this section show that the sug-
gested method with the Kalman observer can provide longitudinal state estimates, in the
absence of road friction details. This method has been experimentally tested with aggres-
sive manoeuvres such as increasing the longitudinal speed during cornering, AiT, harsh
steering, and launch on low-friction surfaces.
 22 E. HASHEMI ET AL.

(a)

(b) (c)
Downloaded by [University of Cambridge] at 00:23 15 June 2016

Figure 12. ûx , BiT and steering on snow/ice.

(a)

(b)

Figure 13. Wheel speed, measured corner velocities, and estimated velocities at wheel centre. for (a)
steering on snow/ice and (b) BiT and Accel. on snow.
 VEHICLE SYSTEM DYNAMICS 23

5. Conclusion
A corner-based longitudinal state estimation robust to the road condition changes is pro-
posed in this article. An UIO and UKF are investigated for estimation of longitudinal forces
and the stability of the UIO for longitudinal force estimation is provided. The important
feature of the developed corner-based force estimator is that it does not implement any
tyre model and is independent from changes in the road friction or tyre parameters due
to wear, inflation pressure, etc. In addition, the suggested force estimators can address the
cases in which tyres are on surfaces with various road frictions (split-μ).
A Kalman-based velocity estimation method using the average lumped LuGre model at
each corner is proposed in this article, and its performance is studied. Taking advantage of
the dynamics on the internal tyre deflection states, the LuGre model facilitates combination
of the vehicle kinematics and tyre forces. Stochastic observability of the proposed Kalman-
based velocity estimator are also investigated. Based on the experimental and simulation
Downloaded by [University of Cambridge] at 00:23 15 June 2016

results, it is concluded that the proposed estimator can handle dry/slippery roads with
pure/combined-slip conditions.
One of significant advantages of the suggested Kalman observer is that the unidirec-
tional lumped LuGre model can be employed instead of the combined-slip model. This
substitution is made possible by denoting the term z(σ0 |Vr |/θ g(Vr )), which contains the
combined friction model, as uncertainty.
The proposed velocity estimation algorithm detects large slip ratio cases with an adap-
tive high-slip threshold (29), based on the excitation level, to allocate adaptive covariance
matrices and tackle the noises associated with harsh manoeuvres. The road experiments
show that the corner-based state estimators can handle dry and slippery roads with the
normalised error RMS ς̄ < 6.25% for the longitudinal force estimation, ς̄ < 3.9% for the
vertical (normal) force, and ς̄ < 6.6% for the longitudinal velocity estimation.
The developed longitudinal estimators can be integrated with active safety systems (e.g.
stability control and traction control) to enhance the performance of such systems in
the presence of road friction changes. Moreover, while preserving the overall structure of
the estimation, one can replace force and velocity estimators independently because of the
modularity of the provided structure.
Observed inconsistencies in some tests of the proposed estimation approach are due to
the lack of a general chassis model, a limitation to be addressed in future works, using the
vehicle generalised force model.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding information
This work was supported by Automotive Partnership Canada, Ontario Research Fund, and the
General Motors Co. [grant numbers APCPJ 395996-09 and ORF-RE-04-039]

References
[1] Ryu J, Gerdes JC. Integrating inertial sensors with global positioning system (GPS) for vehicle
dynamics control. J Dyn Syst Measure Control. 2004;126(2):243–254.
 24 E. HASHEMI ET AL.

[2] Bevly DM, Ryu J, Gerde JC. Integrating INS sensors with GPS measurements for continuous
estimation of vehicle sideslip, roll, and tire cornering stiffness. IEEE Trans Intell Transp Syst.
2006;7(4):483–493.
[3] Imsland L, Johansen TA, Fossen TI, Grip HF, Kalkkuhl JC, Suissa A. Vehicle velocity estimation
using nonlinear observers. Automatica. 2006;42(12):2091–2103.
[4] Imsland L, Grip HF, Johansen TA, Fossen TI, Kalkkuhl JC, Suissa A. Nonlinear observer for
vehicle velocity with friction and road bank angle adaptation–validation and comparison with
an extended Kalman filter. SAE Technical Paper, Tech. Rep.; 2007.
[5] Hsu Y-HJ. Estimation and control of lateral tire forces using steering torque [Ph.D. disserta-
tion]. Stanford University; 2009.
[6] Baffet G, Charara A, Dherbomez G. An observer of tire–road forces and friction for active
security vehicle systems. IEEE/ASME Trans Mechatron. 12(6):651–661.
[7] Hsu Y-HJ, Gerdes JC. The predictive nature of pneumatic trail: tire slip angle and peak force
estimation using steering torque. AVEC08, Kobe, Japan; 2008.
[8] Yoon J-H, Peng H. A cost-effective sideslip estimation method using velocity measurements
from two GPS receivers. IEEE Trans Veh Technol. 2014;63(6):2589–2599.
Downloaded by [University of Cambridge] at 00:23 15 June 2016

[9] Hac A, Simpson MD. Estimation of vehicle side slip angle and yaw rate. SAE Technical Paper,
Tech. Rep.; 2000.
[10] Wenzel TA, Burnham KJ, Blundell MV, Williams RA. Dual extended Kalman filter for vehicle
state and parameter estimation. Veh Syst Dyn. 2006;44(2):153–171.
[11] Julier SJ, Uhlmann JK, Durrant-Whyte HF. A new approach for filtering nonlinear systems.
IEEE proceedings of the 1995 American control conference, Seattle, WA, vol. 3; 1995. p.
1628–1632.
[12] Wan E, Van Der Merwe R, et al. The unscented Kalman filter for nonlinear estimation. In: IEEE
adaptive systems for signal processing, communications, and control symposium (AS-SPCC),
Lake Louise, Alberta; 2000. p. 153–158.
[13] Pasterkamp WR, Pacejka HB. The tyre as a sensor to estimate friction. Veh Syst Dyn.
1997;27(5–6):409–422.
[14] Uil R. Tyre models for steady-state vehicle handling analysis [Ph.D. dissertation]. Eindhoven
University of Technology; 2007.
[15] Antonov S, Fehn A, Kugi A. Unscented Kalman filter for vehicle state estimation. Veh Syst Dyn.
2011;49(9):1497–1520.
[16] Magallan GA, De Angelo CH, Garcia GO, et al. Maximization of the traction forces in a 2wd
electric vehicle. IEEE Trans Veh Technol. 2011;60(2):369–380.
[17] Canudas-De-Wit C, Tsiotras P. Dynamic tire friction models for vehicle traction control.
Proceedings of the 38th conference on decision and control, Phoenix, AZ; 1999. p. 3746–3751.
[18] Zhang X, Xu Y, Pan M, Ren F. A vehicle ABS adaptive sliding-mode control algorithm based
on the vehicle velocity estimation and tyre/road friction coefficient estimations. Veh Syst Dyn.
2014;52(4):475–503.
[19] Burckhardt M. Fahrwerktechnik Radschlupf-regelsysteme. Vogel-Verlag; 1993. p. 16.
[20] Gadola M, Chindamo D, Romano M, Padula F. Development and validation of a Kalman filter-
based model for vehicle slip angle estimation. Veh Syst Dyn. 2014;52(1):68–84.
[21] Canudas-de Wit C, Petersen ML, Shiriaev A. A new nonlinear observer for tire/road distributed
contact friction. Proceedings of the 42nd IEEE conference on decision and control, Maui, HI,
vol. 3; 2003. p. 2246–2251.
[22] Patel N, Edwards C, Spurgeon SK. Tyre–road friction estimation – a comparative study. Proc
Inst Mech Eng Part D: J Autom Eng. 2008;222(12):2337–2351.
[23] Baffet G, Charara A, Lechner D. Estimation of vehicle sideslip, tire force and wheel cornering
stiffness. Control Eng Pract. 2009;17(11):1255–1264.
[24] Hashemi E, Pirani M, Khajepour A, Fidan B, Kasaiezadeh A, Chen S, Litkouhi B. Integrated
estimation structure for the tire friction forces in ground vehicles. in IEEE Conference on
Advanced Intelligent Mechatronics, Banff, Canada; 2016.
[25] Doumiati M, Victorino A, Lechner D, Baffet G, Charara A. Observers for vehicle tyre/road
forces estimation: experimental validation. Veh Syst Dyn. 2010;48(11):1345–1378.
 VEHICLE SYSTEM DYNAMICS 25

[26] Hashemi E, Kasaeizadeh A, Khajepour A, Mushchuk N, Chen S-K. Robust estimation and
experimental evaluation of longitudinal friction forces in ground vehicles. Montreal: ASME
IMECE; 2014.
[27] Mammar S, Glaser S, Netto M. Vehicle lateral dynamics estimation using unknown input
proportional-integral observers. IEEE American control conference, Minneapolis, MN; 2006.
p. 6.
[28] Wang Y, Bevly DM, Chen S-K. Longitudinal tire force estimation with unknown input
observer. ASME 2012 5th annual dynamic systems and control conference. Fort Lauderdale,
FL: American Society of Mechanical Engineers; 2012. p. 523–530.
[29] Hong S, Smith T, Borrelli F, Hedrick JK. Vehicle inertial parameter identification using
extended and unscented Kalman filters. 16th international IEEE conference on intelligent
transportation systems (ITSC), Hague, Netherlands; 2013. p. 1436–1441.
[30] Hong S, Lee C, Borrelli F, Hedrick JK. A novel approach for vehicle inertial parameter
identification using a dual Kalman filter. IEEE Trans Intell Transp Syst. 2015;16(1):151–161.
[31] Haykin Simon E. Kalman filtering and neural networks, vol. 47. New York: John Wiley & Sons
Inc.; 2004.
Downloaded by [University of Cambridge] at 00:23 15 June 2016

[32] Rehm A. Estimation of vehicle roll angle. IEEE 4th international symposium on communica-
tions, control and signal processing (ISCCSP), Limassol, Cyprus; 2010. p. 1–4.
[33] Pacejka HB, Besselink IJM. Magic formula tyre model with transient properties. Veh Syst Dyn.
1997;27(S1):234–249.
[34] Canudas-de Wit C, Tsiotras P, Velenis E, Basset M, Gissinger G. Dynamic friction models for
road/tire longitudinal interaction. Veh Syst Dyn. 2003;39(3):189–226.
[35] Wang Y, Bevly D. Longitudinal vehicle state estimation. General motors summary report, Tech.
Rep.; 2010.
[36] Bryson AE, Ho Y-C. Applied optimal control: optimization, estimation and control. Levittown,
PA: Taylor & Francis; 1975.
[37] Kumar PR, Varaiya P. Stochastic systems: estimation, identification and adaptive control. New
York: Prentice-Hall, Inc.; 1986.
[38] Franklin GF, Powell JD, Workman ML. Digital control of dynamic systems. Menlo Park:
Addison Wesley Longman; 1998.
[39] Anderson BDO, Moore JB. Detectability and stabilizability of time-varying discrete-time linear
systems. SIAM J Control Optim. 1981;19(1):20–32.
[40] Stengel RF. Optimal control and estimation. New York: Courier Corporation; 2012.
[41] Delyon B. A note on uniform observability. IEEE Trans Autom Control. 2001;46(8):1326–1327.
[42] Bageshwar VL, Gebre-Egziabher D, Garrard WL, Georgiou TT. Stochastic observability test for
discrete-time Kalman filters. J Guid Control Dyn. 2009;32(4):1356–1370.
[43] Tóth R. Modeling and identification of linear parameter-varying systems, vol. 403. Berlin
Heidelberg: Springer; 2010.
[44] Anderson BD, Moore JB. Optimal filtering. New York: Dover Publications; 1995.

Appendix
• Detectability and stabilisability:

Definition A.1: The pair [Ak , Ck ] in the linear time-varying discrete-time system with
state update xk+1 = Ak xk + Bk uk and output equation yk = Ck xk is uniformly detectable if
∃0 ≤ c1 ≤ 1, c2 ∈ R+ and q, k2 ≥ 0, such that in case φk1 +q,k1 ϑ ≥ c1 ϑ for some ϑ, k1 ,
then ϑ T V(k1 , k2 )ϑ ≥ c2 ϑ T ϑ, which necessitates the observability grammian V(k1 , k2 ) to be
V(k1 , k2 ) ≥ d1 I > 0 for some d1 [39]


k2
V(k1 , k2 ) = T
φk,k CT C φ ,
1 k k k,k1
(A1)
k=k1
 26 E. HASHEMI ET AL.

where φi,j = φi,i−1 φi−1,j and φi+1,i = Ai as the state transition matrices for i ≥ j. In addition, the
pair [Ak , Bk ] in the linear time-varying discrete-time system (20) without process and measurement
noise effect is stabilisable if ∃0 ≤ c1 ≤ 1, c2 ∈ R+ and q, k2 ≥ 0, such that in case φk2,k2−q ϑ ≥
c1 ϑ for some ϑ, then ϑ T W(k1 , k2 − 1)ϑ ≥ c2 ϑ T ϑ, which necessitates the controllability gram-
mian W(k1 , k2 − 1) to be full rank [44]
2 −1
k
W(k1 , k2 − 1) = φk2 ,k+1 Bk BTk φkT2 ,k+1 . (A2)
k=k1

• Bounded error covariance for the KF: This characteristic for the time-invariant KF has been
proved before, but provided here for convenience. Detectability condition on (A, C) leads to a
linear estimator with matrix K ∗
∗ ∗ ∗ ∗
xk+1 | k = Axk | k−1 + K (yk − Cxk | k−1 ), (A3)
where (A − K ∗ C) is stable. Thus, the error covariance matrix for such estimator is defined by
∗ ∗ ∗
Downloaded by [University of Cambridge] at 00:23 15 June 2016

P̄k+1 |k  E [(xk+1 − x̂k+1 | k )(xk+1 − x̂k+1 | k ) ] that yields

T

∗ ∗ ∗ ∗ T ∗ ∗T
P̄k+1 | k = (A − K C)P̄k (A − K C) + K RK , (A4)
which can be written as
∗ ∗ k+1 ∗
P̄k+1 | k = (A − K C) P̄0 | −1 ((A − K ∗ C)k+1 )T


k
+ (A − K ∗ C)i (K ∗ RK ∗ T + Q)((A − K ∗ C)i )T . (A5)
i=0
The first term vanishes and the second term is also bounded because of the stability of (A −
∗
K ∗ C). Therefore, the error covariance P̄k+1 | k of such linear estimator is bounded. This results in
bounded error covariance P̄k for the Kalman estimator because of the optimality of the KF.
• Defining Mk+1 and Nk+1 for completely uncertain initial covariance/states [42]: The initial
M1 , N1 are attainable by the initial measurement error covariance R0 as T0 0 = R̄0 , which
yields the projector  of a vector onto the orthogonal complement of the range space 
0 = C0T −1
0 ,

0 = I − 0 (0T 0 )∗ 0T ,
X0 X0T = 0 ,
M1 = A0 X0 X0T AT0 ,
N1 = Q̄0 + A0 0 ((0T 0 )∗ )2 0T AT0 , (A6)
where (.)∗ represents pseudo inverse of the matrix (.) and full rank factorisation of 0 is denoted
by X0 . The matrix Mk+1 is then defined using the fresh Ck and the measurement noise Rk as the
following procedure:
Tk k = R̄k + Ck N CkT ,
T T T −1
k = X0,k−1 φk,0 Ck k ,
k = I − k (kT k )∗ KT ,
Xk XkT = k ,
Mk+1 = φk,0 X0,k X0,k
T T
φk,0 , (A7)

in which Xk is the full rank factorisation of k and X0,k  X0 X1 . . . Xk . Employing the condition
λmax (Mf ) = 0 for a finite time, Nk+1 is related to Nk as in Equation (27). In addition, the Sk+1
 VEHICLE SYSTEM DYNAMICS 27

matrix in the simplified Riccati equation P̄k+1 | k = Nk+1 + Sk+1 can be written as
Tk,1
Sk+1 = Ak Sk ATk − Ak Nk CkT −1 + · · · −T
k Ck Nk Ak
T
k ψ
Tk,1
− Ak Nk CkT −1
K I− + · · · −T
k Ck Sk Ak
T
ψ
Tk,1
− Ak Sk CkT −1
K I− + · · · −T
k Ck Nk Ak
T
ψ
Tk,1
− Ak Sk CkT −1
K I− + · · · −T
k Ck Sk Ak .
T
(A8)
ψ
Downloaded by [University of Cambridge] at 00:23 15 June 2016