DOI: 10.48011/asba.v2i1.

1305

Comparison of Nonlinear Receding-Horizon and Extended Kalman Filter

Strategies for Ground Vehicles Longitudinal Slip Estimation
Elias Dias Rossi Lopes*, Gustavo Simão Rodrigues** and Helon Vicente Hultmann Ayala*

*Pontifícia Universidade Católica do Rio de Janeiro, Rio de Janeiro, Brasil

(e-mails: [email protected], [email protected])
** Instituto Militar de Engenharia, Rio de Janeiro, Brasil
(e-mail: [email protected])

Abstract: Friction efforts are present in almost all mechanical applications, due to contact between bodies
and there are many important situations, in which they must be properly controlled. Among these, there
are tire contact forces, which is focus of many studies in autonomous vehicles and control applications on
vehicle systems, since the tire forces and moments are nonlinear and may be modelled as friction efforts.
Any control synthesis focused to optimize its performance must be associated to state estimators, since the
efforts depend on slip variables, as longitudinal slip and sideslip angle, and it is not possible to accurately
measure them. So, in this paper, two state estimation algorithms are evaluated: Extended Kalman Filter
(EKF) and Moving Horizon State Estimation (MHSE), which are applied to a quarter-car model for
longitudinal dynamics. It is presented that, for both traction and braking phases, the MHSE is more
accurate, since it takes explicitly into account the nonlinear model in the estimation process, independently
of Jacobian sensitivities to discontinuities as is the case here. So, it is demonstrated that the developed
estimator may be successfully associated to controllers with the objective of optimize tire performance in
traction and braking control.
Keywords: Extended Kalman Filter, Moving Horizon State Estimation, Tire Dynamics, Nonlinear Efforts,
Ground Vehicles

dependents on tire dynamics, such as braking, traction and

1. INTRODUCTION
stability control. In most applications, the controlled variables,
In many mechanical applications, friction interactions are usually longitudinal slip, attitude angles, sideslip angle and
commonly found, once there is usually contact between others are impossible to measure, which justifies the
bodies. In most cases, these interactions have a nonlinear implementation of observer-based control.
nature, which difficult any attempt of control the system
In this context, Zareian et al. (2015) propose the use of EKF
performance or mitigating their effects, when harmful.
associated to Recursive Least Squares and Neural Networks in
Coulomb friction, tire-road forces in vehicles and bit-rock
a methodology of estimation of road friction coefficient, which
interaction on perforation process are some examples of these
is a hardly obtaining parameter. Kayacan et al. (2018) present
effects.
a control strategy for tracked field robots with receding
However, in a control application, it is not common to be horizon estimation and control (RHEC). The estimation
possible to measure all states or all controlled variables. In algorithm is used for estimate states and parameters, and the
these situations, it is usually used a state observer, which has receding horizon control is based on an adaptive system whom
as function estimate the states based on the measured model is time varying. Li et al. (2014) present an EKF based
variables. Among the state estimation algorithms, the Kalmar estimator for sideslip angle for a vehicle stability control and
is one of the most known, such as one of its versions applied the authors remark that its measure is complex and expensive,
to nonlinear systems, the Extended Kalman Filter (EKF). On which justifies the estimation process. Sun et al. (2014) apply
last decades, other algorithms have been developed, looking a nonlinear observer for state estimation on an ABS, due to
for more robustness and accuracy. The Moving Horizon nonlinearity of the friction force during brake.
Estimation (MHE) is one of these, defined by Alessandri et al.
Boada et al. (2017) develop a new method for estimation of
(2008) as a powerful and robust approach, suitable in systems
different states and parameters of a vehicle, using a
with modelling uncertainties and numerical errors. It was
constrained version of Kalman Filter to consider the physical
developed as a dual of Model Predictive Control and estimates
limitations of the parameters. It is demonstrated by
the states variables using a defined horizon of recent
experimental results that the constraints are important to
information and measures. These algorithms are studied in
improve accuracy of the algorithm. Na et al. (2018) present
many different applications (Brembeck, 2019), so that it is
two torque estimation methods for vehicle engines, using a
possible to clearly understand its power and robustness.
proper dynamical model and air mass flow rate and engine
In vehicle systems, nonlinear observers are implemented on speed, which are measurable. Jo et al. (2016) present a road
many control applications, but we may specially remark those slope and position estimator, which inputs are GPS data and
vehicle onboard sensors. The estimator proposed presents remarks that state-estimators for autonomous vehicles are even
more accurate and reliable results, which is proven by more challenging, since the complexity of models and
experiments. Hsiao (2012) proposes an observer-based control applications rises along the time. In this way, he presents a
scheme for traction force, robust to variations on road vehicle state observer to estimate position, yaw angle and their
conditions and uncertainties on tire models. rates, with focus on path following and he discuss about the
balance between model complexity and estimator
Nilsson et al. (2014) study the problem of estimating position performance. The author uses constrained versions of EKF and
and direction of a vehicle with a single camera since it is hardly MHSE to better approximate the results to real data. Jalali et
dependent of image quality. So, the authors propose an al. (2017) present a model predictive control scheme for
estimator which combines onboard vehicle sensors and tracking yaw rate with small lateral velocity and tire slips. The
adjusted camera images, with a single-track model. Chen et al. proposed method controls lateral velocity adjusting reference
(2011) remark the importance of tire-road friction coefficient yaw rate, which reduces the size of model and computational
estimation for autonomous vehicle applications and present an complexity. They also present an estimation algorithm by
observer which does not depend on longitudinal motion means of vehicle kinematics and tire model.
information and is properly associated to an adaptive speed
control. Singh et al. (2013) remark that simpler stability The state estimation is possible only if the system is fully
control performs well in many situations, but it is improved observable. The condition of observability of a system is
when a tire-road friction estimator is associated to the control characterized by the possibility of observe all state variables
scheme. In this way, the authors present a method in which is by means of the measurement variables, or yet, if two different
used frequency response of tire vibrations on the estimation sets of states are related to two different sets of measured
algorithm. Hsu et al. (2009) remark the importance of variables (Kou et al., 1973). The authors explore the
knowledge of physical limits of parameters used on vehicle observability of nonlinear systems, presenting two sufficient
control, such as tire slip angle and maximum lateral force and conditions to prove it. Katriniok and Abel (2015) present an
propose a model-based estimation algorithm that estimates EKF estimation for longitudinal and lateral velocities and yaw
them using information from the applied steering torque. rate. They also present an approach for evaluate local
observability online and a virtual measurement variable for
Du et al. (2015) construct a side-slip estimator based on a instants in which local observability is lost.
fuzzy system for lateral dynamics and the nonlinear Dugoff
tire model, using measured yaw rate and estimated states. Li et In this context, this work aims to present a comparative
al. (2019) use the same tire model to propose a side-slip analysis between EKF and MHSE for estimation of
estimation algorithm robust to inaccurate tire parameters. longitudinal slip in ground vehicle control applications. We
may observe on the literature review that the MHSE is not
Recent research in vehicle control point the increased use of exploited on vehicle systems, and thus the main goal of the
electric in-wheel motors, which allows many control strategies paper is to assess its performance in such applications. Both
and simpler configurations of electric vehicles. These devices estimators are applied in the same conditions to estimate states
allow to reduce mass and to simplify transmission system, of longitudinal dynamics of a quarter-car model, actuated by a
which is favorable in electric and autonomous vehicles. Zhao braking or traction torque in the case of an in-wheel motor. In
and Liu (2014) present a four degree-of-freedom nonlinear the estimation process, the state estimation algorithms are
dynamical model of a four independent wheel electric vehicle, employed to estimate accurately the longitudinal slip and
considering the measurements provided by modern sensors velocity of the vehicle. We observe in the literature review that
used on vehicles. An observer is associated to this model to the comparison has not been assessed thus far in such
estimate vehicle velocity and roll angle, since these variables application, even if results in the literature show overall better
must be controlled on stability control system. Feng et al. results for the receding-horizon approaches (Alessandri et al.,
(2020) present two estimation algorithms based on moving 2008). It is important to remark that the longitudinal slip is
horizon methods. The observer is applied on a four wheels important for traction and braking control strategies. The
electric robotic platform on different friction conditions. present paper aims at the evaluation of the performance of
Jeon et al. (2018) propose a real-time constrained Kalman MHSE and EKF, their limitations and advantages, aiming at
filter algorithm for estimation of the three tire forces on vehicle future control application on autonomous vehicles. The
tires, namely, vertical, longitudinal and lateral forces in mobile present work shows overall favourable results for the MHSE
robots equipped with wheel encoders and navigation sensors. approach, despite its greater computational effort.
Tire forces in the estimation process are modeled by Magic At first, it is presented the theoretical basis of quarter-car
Formula, an empirical model developed by Pacejka (1992). longitudinal dynamic model, studying the longitudinal force
Hong et al. (2014) present an application of Unscented Kalman formulations and the parameters employed. In the following
Filter to estimation of inertial parameters of vehicles, which section, the estimation processes, namely, the Extended
may be not accurately determined in design phase. Heidfeld et Kalman Filter and the Moving Horizon State Estimation are
al. (2019) applied the same algorithm in a state and tire slip defined, and their algorithms and evaluation metrics are
estimation for an electric vehicle with four independent presented. Then, it is demonstrated the data obtaining by
wheels. means of a simulation with noisy measured variables and the
Estimation algorithms are used also in path-tracking results of the estimation process. Finally, the conclusions are
applications for autonomous vehicles. Brembeck (2019)
commented, evaluating both estimators, and presenting the parameters, which are empirically obtained. The last one has
possibilities of future research. the advantage of being continuous in whole domain of slip,
which does not happen on the others, reducing elapsed time of
2. QUARTER-CAR DYNAMIC MODEL simulations and estimation processes. So, according to the
Magic Formula model, µ may be written as:
A quarter-car model may be effectively used for the study of
nonlinear estimation. In this model, the vehicle is understood 𝜇(𝜆) = 𝐴 sin(𝐵 atan(𝐶𝜆 − 𝐷(𝐶𝜆 − atan(𝐶𝜆)))) (6)
as a concentrated mass (with mass m) over a single wheel (with
moment of inertia J), and there are no effects related to vertical The parameters A, B, C and D depends on road conditions on
or lateral dynamics. Are considered also the rolling resistance which the vehicle moves. Figure 2 presents the curves
momentum and aerodynamic resistance (Figure 1). associated to many road conditions, which are obtained by
approximation to data presented by Savaresi (2005).
The first one affects the wheel dynamics and it is due to energy
dissipation in the tire structure and rubber (Jazar, 2017).
Mathematically, it may be described as proportional to normal
load, according to the coefficient fR. The drag force acts on
vehicle gravity center and is due to aerodynamic efforts. For
simplification, we must assume that it is proportional to the
square of longitudinal velocity according to C. So, the
dynamic equations of the system are:
𝑚𝑣̇ = 𝐹𝑥 − 𝑅𝐴𝑒𝑟 = 𝐹𝑥 − 𝐶𝑣 2 (1)
𝐽𝜔̇ = 𝑇 − 𝑟𝐹𝑥 − 𝑀𝑅𝑜𝑙 = 𝑇 − 𝑟𝐹𝑥 − 𝑟𝑓𝑅 𝑚𝑔 (2)

Figure 2: Friction coefficients for different pavements.

Furthermore, we must define the longitudinal slip and the
velocity as states of the system, since they are controlled
variables on many applications, and the tire angular velocity
as the output variable, since it is usually measured. In this way,
from equation (5), the ω must be written in terms of v and λ
(equation (7)) and we may obtain its time derivative (equation
(8)).
𝑣(1 + 𝜆)
𝜔= (7)
Figure 1: Schematic model of quarter car 𝑟
𝑣̇ (1 + 𝜆) + 𝜆̇𝑣
In the wheel dynamic equation, T is the torque, which is 𝜔̇ = (8)
considered the system input. The traction force Fx is defined 𝑟
as proportional to the normal load (in the quartel-car model Generally, we may define a nonlinear state space model as:
defined as the weight) according to a factor µ (Savaresi, 2005).
Then, treating c = C/m, the motion equations may be written 𝒙̇ = 𝑓(𝒙, 𝒖)
{ (9)
as: 𝒛 = ℎ(𝒙, 𝒖) + 𝝃𝑛
𝑣̇ = 𝜇(𝜆)𝑔 − 𝑐𝑣 2 (3) In equation (9), x is the vector of states, u the input of the
system, z are the outputs, that is, the measured states and 𝝃𝒏 is
𝑇 𝑚𝑔𝑟
𝜔̇ = − (𝜇(𝜆) + 𝑓𝑅 ) (4) a white measurement noise, which may be considered on
𝐽 𝐽 simulation process and data obtaining. Substituting equation
The friction coefficient µ depends on the longitudinal slip λ, (8) on (4), we obtain the nonlinear state space equations of the
which may be define as: system.

𝜔𝑟 − 𝑣 𝜔𝑟 𝑇𝑟 (1 + 𝜆) 𝑚𝑔𝑟 2
𝜆= = −1 (5) 𝜆̇ = − (𝜇(𝜆)𝑔 − 𝑐𝑣 2 ) − (𝜇(𝜆) + 𝑓𝑅 ) (10)
𝑣 𝑣 𝐽𝑣 𝑣 𝐽𝑣

We may note that during an acceleration, slip is positive, and 𝑣̇ = 𝜇(𝜆)𝑔 − 𝑐𝑣 2 (11)
during braking, it is negative and equation (5) demonstrate that 𝑣(1 + 𝜆)
λ must be on [-1,∞[. There are many formulations for the 𝜔= (12)
𝑟
relationship between µ and λ, as the Julien Theory (Lopes et
al., 2019), the Burckhardt model (Savaresi, 2005) and the For Kalman filter application, it was remarked that the
Magic Formula (Pacejka, 1992). All of them depends on many Jacobians of the nonlinear model should be defined. These
Jacobians correspond, respectively, to state and output 3. Residual covariance
matrices of a linear state space model. For longitudinal
dynamics, they are: 𝑆(𝑖 + 1) = 𝑅(𝑖) + 𝐻(𝑖)𝑃(𝑖 + 1|𝑖)𝐻(𝑖)𝑇 (20)

𝜕𝜆̇ 𝜕 𝜆̇ 4. Filter gain

𝜕𝒇
𝑭= = 𝜕𝜆 𝜕𝑣 (13) 𝑊(𝑖 + 1) = 𝑃(𝑖 + 1|𝑖)𝐻(𝑖)𝑇 𝑆(𝑖 + 1)−1 (21)
𝜕𝒙 𝜕 𝑣̇ 𝜕 𝑣̇
[ 𝜕𝜆 𝜕𝑣 ] 5. State prediction
𝜕𝒉 𝜕𝜔 𝜕𝜔 𝑣 1+𝜆
𝑯= =[ ]=[ ] (14) 𝑥̂(𝑖 + 1|𝑖) = 𝑓(𝑥̂(𝑖|𝑖), 𝑢(𝑖)) (22)
𝜕𝒙 𝜕𝜆 𝜕𝑣 𝑟 𝑟
6. Measurement prediction
As mentioned on the previous section, the estimation of all
states is possible if, and only if, the system is full observable. 𝑧̂ (𝑖 + 1|𝑖) = ℎ(𝑥̂(𝑖 + 1|𝑖)) (23)
Specifically on the quarter car model, we may calculate the
observability matrix based on the Jacobians of the nonlinear 7. Measurement residual
model, with the classical observability matrix, defined as:
𝜈(𝑖 + 1) = 𝑧(𝑖 + 1) − 𝑧̂ (𝑖 + 1|𝑖) (24)
𝐻
𝒪=[ ] (15) 8. Updated state estimate
𝐻𝐹
On the simulation process, it is possible to note that the 𝑥̂(𝑖 + 1|𝑖 + 1) = 𝑥̂(𝑖 + 1|𝑖) + 𝑊(𝑖 + 1)𝜈(𝑖 + 1) (25)
observability matrix has full rank in all time. Consequently, all
states are observables and, in a first analysis, both estimation 9. Updated state covariance
algorithms may be successfully employed. 𝑃(𝑖 + 1|𝑖 + 1)
(26)
3. STATE ESTIMATION METHODS = 𝑃(𝑖 + 1|𝑖) − 𝑊(𝑖 + 1)𝑆(𝑖 + 1)𝑊(𝑖 + 1)𝑇

In a control application, it is not possible to always measure all The MHSE is most recent and have its development related to
states of the system. In these cases, an estimator must be Model Predictive Control, since there is a duality between
defined, with the objective of estimate all states in each instant, regulation and estimation processes (Alessandri, 2008). In this
based on output variables, that is, the measured ones. The method, the estimation of states in each instant of time is
existence of an estimator is conditioned to the observability of obtained by means a prediction of the states N instants before
the system, which indicates that all states may be observed by and the estimation of the states in this window of time using
means of the output variables. the dynamic model and the measured data. In this case,
continuous or discrete-time models may be used. The
In this work, two methods are explored: Extended Kalman algorithm of MHSE is:
Filter (EKF) and the Moving Horizon State Estimation
(MHSE). The Extended Kalman Filter is one of nonlinear 1. Prediction of states at t-N
applications of the Kalmar Filter, developed for linear systems
𝑥̅ (𝑡 − 𝑁|𝑡)
(Bar-Shalom et al., 2004). (27)
= 𝑓(𝑥̂(𝑡 − 𝑁 − 1|𝑡 − 1), 𝑢(𝑡 − 𝑁 − 1))
The EKF is applied to a discrete-time system, such as:
2. State estimation at t-N
𝒙(𝑖 + 1) = 𝒇(𝒙(𝑖), 𝒖(𝑖))
{ (16) 𝑥̂(𝑡 − 𝑁|𝑡)
𝑧(𝑖 + 1) = 𝒉(𝒙(𝑖 + 1))
= 𝑎𝑟𝑔𝑚𝑖𝑛 (𝜇‖𝑥̂(𝑡 − 𝑁|𝑡) − 𝑥̅ (𝑡 − 𝑁|𝑡)‖
Briefly, the algorithm for EKF is described on the sequence (28)
below, for a state estimate (𝑥̂(𝑖|𝑖)), that is the system state at 𝑡
sample i, estimated on sample i. It is important to remark that + ∑ ‖ℎ(𝑥̂(𝑖|𝑡)) − 𝑧(𝑖)‖)
the matrices P, R and Q must be initialized as diagonal types, 𝑖=𝑡−𝑁
with large traces, to assure the convergence of the filter.
3. State estimation at the horizon
1. Jacobians:
𝑥̂(𝑖 + 1|𝑡) = 𝑓(𝑥̂(𝑖|𝑖), 𝑢(𝑖)) 𝑖 = 𝑡 − 𝑁, … , 𝑡 − 1 (29)
𝜕𝑓
𝐹(𝑖) = (17) 4. State estimation at t
𝜕𝑥 𝑥=𝑥̂(𝑖 |𝑖 )

𝜕ℎ 𝑥̂(𝑡|𝑡) = 𝑓(𝑥̂(𝑡 − 1|𝑡), 𝑢(𝑡 − 1)) (30)

𝐻(𝑖) = (18)
𝜕𝑥 𝑥=𝑥̂(𝑖 |𝑖 ) On the algorithm, it is important to remark some observations.
The first one is related to the parameter µ, which indicates the
2. State prediction covariance:
confidence on the state prediction, that is, with this parameter,
𝑃(𝑖 + 1|𝑖) = 𝐹(𝑖)𝑃(𝑖|𝑖)𝐹(𝑖)𝑇 + 𝑄(𝑖) (19) it may be differently considered the dynamic model and state
prediction or the measured data, on the cost function. The
second observation is related to the optimization process, zero, since the Jacobians have discontinuities. For MHSE, in
which must be defined as so fast it is possible, since, in a this situation, the RMSE is equal zero.
control application, the estimation must occur into the time
between two samples. In a second scenario, in which it is supposed that the
measurement noise (𝝃𝑛 ) has normal distribution with standard
The evaluation of both methods may be done with some deviation of 1%, we have the results presented on the Figure 4
metrics (Alessandri et al., 2008). One of them is the Root Mean and Figure 5. In this situation, the input torque previously
Square Error, which defines if the variables are quite equal in presented is applied on the wheel. It is possible to observe that
all time instants, for n simulations. It is defined also the it is suitable to reach a constant velocity, even with a steady-
Asymptotic Root Mean Square Error (ARMSE), which state error, which is no focus on this paper. The most important
measure the RMSE on the final window of time S, considering result we may remark is that the non-measured variables are
T the final simulation time. estimated with good accuracy.
1⁄
𝑛 2
‖𝑒(𝑡, 𝑖)‖2
𝑅𝑀𝑆𝐸(𝑡) = (∑ ) (31)
𝑛
𝑖=1
1⁄
𝑇 𝑛 2
1 ‖𝑒(𝑡, 𝑖)‖2
𝐴𝑅𝑀𝑆𝐸 = ∑ (∑ ) (32)
𝑆+1 𝑛
𝑇−𝑆 𝑖=1

4. SIMULATION AND STATE ESTIMATION

The quarter-car model is simulated with parameters of a
typical passenger vehicle, moving on dry asphalt. To evaluate
the system and the estimators, its dynamics is simulated
considering an input torque defined with a proportional control
law, so that the vehicle reaches a reference speed of 20 m/s. Figure 4: Estimated longitudinal velocity (m/s)
Then:

𝑇 = 𝑘𝑝 (𝑣𝑟𝑒𝑓 − 𝑣) (33)

Considering this condition, the input torque considered on the

simulated system is presented on Figure 3.

Figure 5: Estimated wheel rotation (rad/s)

For the simulation above, in which the initial conditions are
well-known and the wheel rotation is measured with a noise of
Figure 3: Input torque (Nm) 1%, RMSE is presented on Table 1.
On the simulations, it is considered that output variable, that is Table 1: RMSE results with well-known initial conditions
the wheel rotation, is measured with different amplitude noises
(𝝃𝑛 ). So, the main objective of the estimators is to obtain the State EKF MHSE Reduction
states along the time, based on this measured one. Evidently,
as the system is full observable, when no noise is considered λ 0.0615 0.0025 95.935 %
on the measurement devices, the states are precisely estimated.
𝑣 0.0210 0.0023 89.048 %
We must remark that, for EKF, this precision depends on the
sample time adopted on discretization process and are close to
We must do two remarks about these results. The first one is random numbers on interval [0,1], many times. This way, it is
related to slip estimation. As there is no relation between the possible to observe its tendency in more sampled data and to
definition of input torque and slip, the initial values of verify if the estimators are capable to correct states even if the
longitudinal slip are very high, which represents the situation initial conditions are badly defined or unknown.
of the wheel slipping on the road. This may be observed on
Figure 4, since the wheel rotation rises so quickly. In a second Table 2 presents the evaluation of the mean and the standard
phase, the rotation falls, reducing the longitudinal slip to deviation of the ARMSE for longitudinal slip (λ), which is one
values between 0 and 1. The second remark is related to the of the non-measured states and is used on control strategies. In
Jacobians. The state equation related to the longitudinal slip this table are presented the simulation results for different
presents many discontinuities points, mainly related to noise standard deviations, after the convergence of RMSE(t),
situations in which the velocity is zero. In this way, when this briefly defined as ARMSE (equation (32)).
state is very low, the Jacobian F presents high values, harming
the estimation process and also compromising the Table 2: ARMSE analysis for longitudinal slip (λ) estimation
convergence. To prevent this situation, the same constraints
Algorithm EKF MHSE
applied to state variables are applied during estimation
process. As the MHSE does not depend on derivatives, it is not Mean 0.0084 4.8188e-5
affected by this situation and, so, its results are much better, 𝝃𝑛 = 0.001
presenting high relative reductions on RMSE. Figure 6 St. Dev. 4.5271e-5 2.4763e-5
presents the estimated longitudinal slip in a converged
simulation. It is important to remark that the states are Mean 0.4275 5.1430e-4
estimated based on a noisy signal and compared with the 𝝃𝑛 = 0.01
St. Dev. 0.0043 2.9450e-4
supposed real one, which is noiseless. It is remarkable also the
nearest values presented by the MSHE result (Figure 7). Mean 2.7254 0.0027
𝝃𝑛 = 0.05
St. Dev. 0.0363 0.0016

A more detailed analysis may be done interpreting the

RMSE(t), as defined by equation (31), and that is presented on
Figure 8, Figure 9 and Figure 10 for, respectively, standard
deviations of 0.001, 0.01 and 0.05 on the measure noise.

Figure 6: Estimated longitudinal slip.

Figure 8: RMSE(t) for measurement noise of 0.1%

On the first scenario, it may be noted that in both methods the
values of errors are larger in the initial samples, because it is
supposed that the initial conditions are unknown and are
treated as different to zero, when they really have this value on
the simulation for data obtaining. Then, the RMSEs fall and
converge for both algorithms, but for the MHSE, the value is
lower than for EKF, denoting that the first one has a better
Figure 7: Estimated longitudinal slip, in detail. accuracy. Besides that, the MHSE converges more quickly
In other analysis, we may realize the estimation process with than EKF. In this way, we may affirm that the MHSE corrects
different initial conditions, defined by uniformly distributed badly defined initial conditions better than EKF. The last one
presents an irregular behavior, with high variation before performance of MHSE on nonlinear mechanical systems with
converging and even the final value is higher than MHSE one. discontinuities, as friction efforts.

Table 3: ARMSE analysis for velocity (v) estimation

Algorithm EKF MHSE

Mean 0.9756 0.0014

𝝃𝑛 = 0.001
St. Dev. 0.0041 7.2911e-4

Mean 3.8729 0.0147

𝝃𝑛 = 0.01
St. Dev. 0.0154 0.0083

Mean 14.3821 0.0788

𝝃𝑛 = 0.05
St. Dev. 0.0491 0.0478

Figure 9: RMSE(t) for measurement noise of 1% 5. CONCLUSIONS

We may conclude about the efficiency of the MHSE in
On the second scenario, the convergence occurs in the same comparison to EKF on state estimation for nonlinear
configuration, that is, both converges, but MHSE converge to applications, especially with friction and discontinue efforts.
lower values in shorter times compared to EKF. Besides that, In the studied case, both algorithms estimate the non-measured
we may observe that the converged value of EKF is higher than states. However, in a performance analysis, the MHSE
the first case. The same aspect is observed on the last scenario presents better results, especially when it is considered the
(Figure 10). measurement errors and uncertainties in initial conditions of
the system.
In future works, it is suggested the application of the
estimation algorithm in complete vehicle models, considering
all wheels and its lateral dynamics, in which it is possible to
evaluate the estimator behavior on different maneuvers. In this
sense, all tire longitudinal slip must be accurately estimated,
which possibilities suitable control strategies for agile and
high-speed path tracking.
Other future possible work is the definition of more complex
control strategies with the objective of achieving prescribed
velocities with optimized longitudinal slip. We must remark
that in autonomous vehicle applications, state observers allow
the possibility to define suitable controller for path tracking,
agility, or stability.
It is also suggested the application of the developed estimators
in other mechanical applications in which the friction efforts
Figure 10: RMSE(t) for measurement noise of 5%
must be mitigated or estimated properly, especially on
discontinuous systems, as, for example, with Coulomb
For the velocity estimation, the results are presented on Table friction, in which MHSE must present better performance
3. The results are similar to the longitudinal slip, which also compared to Kalman Filter approaches, since it is robustness
demonstrates that MHSE has better performance in demonstrated on the results section.
comparison to EKF.
REFERENCES
In a general way, EKF do not present good results, when
compared to MHSE in nonlinear mechanical systems. It may Alessandri, A., Baglietto, M., & Battistelli, G. (2008).
be noted that EKF is strongly affected by measurement noises, Moving-horizon state estimation for nonlinear discrete-
presenting a RMSE on the converged region higher than time systems: New stability results and approximation
MHSE, mainly due to discontinuities on state equations. The schemes. Automatica, 44(7), 1753-1765.
second method is more robust to discontinuities on Jacobians Bar-Shalom, Y., Li, X. R., & Kirubarajan, T.
and state equation, and the errors and time to convergence (2004). Estimation with applications to tracking and
presented are lower. So, we may conclude about the higher
 navigation: theory algorithms and software. John Wiley observability. IEEE Transactions on Control Systems
& Sons. Technology, 24(4), 1368-1381.
Boada, B. L., Garcia-Pozuelo, D., Boada, M. J. L., & Diaz, V. Kayacan, E., Young, S. N., Peschel, J. M., & Chowdhary, G.
(2016). A constrained dual Kalman filter based on pdf (2018). High‐precision control of tracked field robots in
truncation for estimation of vehicle parameters and road the presence of unknown traction coefficients. Journal of
bank angle: Analysis and experimental validation. IEEE Field Robotics, 35(7), 1050-1062.
Transactions on Intelligent Transportation Kou, S. R., Elliott, D. L., & Tarn, T. J. (1973). Observability
Systems, 18(4), 1006-1016. of nonlinear systems. Information and Control, 22(1), 89-
Brembeck, J. (2019). Nonlinear constrained moving horizon 99.
estimation applied to vehicle position Li, B., Du, H., Li, W., & Zhang, B. (2019). Non-linear tyre
estimation. Sensors, 19(10), 2276. model–based non-singular terminal sliding mode observer
Chen, Y., & Wang, J. (2011). Adaptive vehicle speed control for vehicle velocity and side-slip angle
with input injections for longitudinal motion independent estimation. Proceedings of the Institution of Mechanical
road frictional condition estimation. IEEE Transactions Engineers, Part D: Journal of automobile
on Vehicular Technology, 60(3), 839-848. engineering, 233(1), 38-54.
Du, H., Lam, J., Cheung, K. C., Li, W., & Zhang, N. (2015). Li, L., Jia, G., Ran, X., Song, J., & Wu, K. (2014). A variable
Side-slip angle estimation and stability control for a structure extended Kalman filter for vehicle sideslip angle
vehicle with a non-linear tyre model and a varying estimation on a low friction road. Vehicle System
speed. Proceedings of the Institution of Mechanical Dynamics, 52(2), 280-308.
Engineers, Part D: Journal of Automobile Lopes, E. D. R., Pinto, A. F. A., Valentim, M. X. G., Peixoto,
Engineering, 229(4), 486-505. P. S., & Neto, R. T. D. C. (2019). Extended model for
Feng, Y., Chen, H., Zhao, H., & Zhou, H. (2020). Road tire calculation of soil-wheel contact area parameters in rigid
friction coefficient estimation for four wheel drive electric soil-deformable tyre approximation. International
vehicle based on moving optimal estimation Journal of Vehicle Systems Modelling and Testing, 13(4),
strategy. Mechanical Systems and Signal 358-372.
Processing, 139, 106416. Na, J., Chen, A. S., Herrmann, G., Burke, R., & Brace, C.
Heidfeld, H., Schünemann, M., & Kasper, R. (2019). UKF- (2017). Vehicle engine torque estimation via unknown
based State and tire slip estimation for a 4WD electric input observer and adaptive parameter estimation. IEEE
vehicle. Vehicle System Dynamics, 1-18. Transactions on Vehicular Technology, 67(1), 409-422.
Hong, S., Lee, C., Borrelli, F., & Hedrick, J. K. (2014). A Nilsson, J., Fredriksson, J., & Ödblom, A. C. (2014). Reliable
novel approach for vehicle inertial parameter vehicle pose estimation using vision and a single-track
identification using a dual Kalman filter. IEEE model. IEEE Transactions on Intelligent Transportation
Transactions on Intelligent Transportation Systems, 15(6), 2630-2643.
Systems, 16(1), 151-161. Pacejka, H. B., & Bakker, E. (1992). The magic formula tyre
Hsiao, T. (2012). Robust estimation and control of tire traction model. Vehicle system dynamics, 21(S1), 1-18.
forces. IEEE transactions on vehicular technology, 62(3), Savaresi, S. M., Tanelli, M., Cantoni, C., Charalambakis, D.,
1378-1383. Previdi, F., & Bittanti, S. (2005). Slip-deceleration control
Hsu, Y. H. J., Laws, S. M., & Gerdes, J. C. (2009). Estimation in anti-lock braking systems. IFAC Proceedings
of tire slip angle and friction limits using steering Volumes, 38(1), 103-108.
torque. IEEE Transactions on Control Systems Singh, K. B., Ali Arat, M., & Taheri, S. (2013). An intelligent
Technology, 18(4), 896-907. tire based tire-road friction estimation technique and
Jalali, M., Hashemi, E., Khajepour, A., Chen, S. K., & adaptive wheel slip controller for antilock brake
Litkouhi, B. (2017). Integrated model predictive control system. Journal of Dynamic Systems, Measurement, and
and velocity estimation of electric Control, 135(3).
vehicles. Mechatronics, 46, 84-100. Sun, F., Lolenko, K., & Rudolph, J. (2014). Nonlinear
Jazar, R. N. (2017). Driveline Dynamics. In Vehicle observer design for state estimation during antilock
Dynamics (pp. 173-223). Springer, Cham. braking. Proceedings of the Institution of Mechanical
Jeon, S. Y., Chung, R., & Lee, D. (2018, October). Tire Force Engineers, Part I: Journal of Systems and Control
Estimation of Dynamic Wheeled Mobile Robots using Engineering, 228(2), 78-86.
Tire-Model Based Constrained Kalman Filtering. In 2018 Zareian, A., Azadi, S., & Kazemi, R. (2016). Estimation of
IEEE/RSJ International Conference on Intelligent Robots road friction coefficient using extended Kalman filter,
and Systems (IROS) (pp. 2470-2477). IEEE. recursive least square, and neural network. Proceedings of
Jo, K., Lee, M., & Sunwoo, M. (2015). Road slope aided the Institution of Mechanical Engineers, Part K: Journal
vehicle position estimation system based on sensor fusion of Multi-body Dynamics, 230(1), 52-68.
of GPS and automotive onboard sensors. IEEE Zhao, L., & Liu, Z. (2014). Vehicle velocity and roll angle
Transactions on Intelligent Transportation estimation with road and friction adaptation for four-
Systems, 17(1), 250-263. wheel independent drive electric vehicle. Mathematical
Katriniok, A., & Abel, D. (2015). Adaptive EKF-based vehicle Problems in Engineering, 2014.
state estimation with online assessment of local