Estimation of Tire-Road Friction for Autonomous Vehicles: a Neural Network Approach

by   Alexandre M. Ribeiro, et al.
University of Campinas

The performance of vehicle active safety systems is dependent on the friction force arising from the contact of tires and the road surface. Therefore, an adequate knowledge of the tire-road friction coefficient is of great importance to achieve a good performance of different vehicle control systems. This paper deals with the tire-road friction coefficient estimation problem through the knowledge of lateral tire force. A time delay neural network (TDNN) is adopted for the proposed estimation design. The TDNN aims at detecting road friction coefficient under lateral force excitations avoiding the use of standard mathematical tire models, which may provide more efficient methods and more robust results. Moreover, the approach is able to estimate the road friction at each wheel independently, instead of using lumped axle models simplifications. Simulations based on a realistic vehicle model are carried out on different road surfaces and driving maneuvers to verify the effectiveness of the proposed estimation method. The results are compared with a classical approach, a model-based method modeled as a nonlinear regression.



There are no comments yet.


page 1

page 2

page 3

page 4


Assessment of Deep Convolutional Neural Networks for Road Surface Classification

When parameterizing vehicle control algorithms for stability or trajecto...

Estimation and Decomposition of Rack Force for Driving on Uneven Roads

The force transmitted from the front tires to the steering rack of a veh...

Lateral Force Prediction using Gaussian Process Regression for Intelligent Tire Systems

Understanding the dynamic behavior of tires and their interactions with ...

Transfer Learning for Input Estimation of Vehicle Systems

This study proposes a learning-based method with domain adaptability for...

Model Guided Road Intersection Classification

Understanding complex scenarios from in-vehicle cameras is essential for...

Abnormal Road Surface Detection Using Wheel Sensor Data

In this manuscript we demonstrate that accurate road abnormality detecti...

Detecting Unsigned Physical Road Incidents from Driver-View Images

Safety on roads is of uttermost importance, especially in the context of...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

One of the primary challenges of vehicle control is that the source of force generation is strongly limited by the available friction between the tire tread elements and the road. In order to better understand vehicle handling due to force generation mechanisms, several research activities related to vehicle dynamics and control fields are oriented towards estimation of all components of the tire-ground contact.

The knowledge of specific tire-road contact operation points, such as the saturation point where the tire generates the maximum force available from friction, could lead to a new range of applications in vehicle control. Moreover, current commercial vehicle safety systems, such as Anti-lock Brake Systems (ABS), Traction Control Systems (TCS) and Electronic Stability Control (ESC) could have a significant improvement in performance by the knowledge of the full vehicle states and operating conditions that are still limited by the lack of information singh2014Estimation . Therefore, to reach this full potential the recognition of the tire’s limit handling is indispensable. With this is mind, we emphasize the importance of estimating the vehicle road conditions, specifically the Tire Road Friction Coefficient (TRFC).

Friction estimation often relies on a model-based estimator using a well-defined and interpretable mathematical model for the purpose of capturing the inherent friction effects under the tire dynamics (forces and moments). The general most common model-based approaches use the steering system model

Liang2019 ; hsu2006method , quarter-car model AlZughaibi2018 , four-wheel vehicle dynamic model ahn2011robust , powertrain and wheel dynamics model Castro2010 .

For the estimation problem, in pasterkamp1997 the correlation between the self-aligning moment of the steering wheels and road friction was firstly examined. Lately, hsu2009estimation ; hsuJournal ; ahn2011robust ; ahnJournal2013 explored the use of a nonlinear Recursive Least Squares method employed as a mean for identification of tire-road friction through an observed data composed by the self-aligning moment and also expanded for a lateral dynamic force analysis.

The wheel dynamic model can also be utilized with a tire model to estimate the friction. In Muller2003 ; Hsiao2011 ; Rajamani2012 the wheel rolling motion is used to detect the longitudinal force and longitudinal friction adopting the powertrain configuration and wheel drive engine. The estimator is built primarily exploring the force-slip ratio plane and its relationship with the road friction coefficient.

Another model-based approach discussed in literature is a slip-slop algorithm. This method is based on the assumption that the low slip-plane zone (linear region of the force-slip plane, characterizing normal driving conditions) can be used to estimate the tire-road friction. Distinct studies Li2007 ; Qi2015 ; Xia2016 have shown this methodology.

Despite the majority of model-based methods, a number of algorithms have been studied based on different concepts to estimate the surface condition. In Casselgren2007 ; Tuononen2008 , an optical sensor is used as a tire sensor that can measure the road ahead and the tire carcass deflections which may be exploited in the estimation of friction potential. Cameras are also used to identify different surfaces. The detection is based on the light polarization changing when reflected from road surface Jokela2009 . Also, Jonsson2011 proposed a method that merges weather data and road images taken by a camera on the vehicle. More recently, based on the hypothesis that the friction coefficient affects the natural frequency of the vehicle systems, such as in-wheel motor drive system or steering system, the road-friction is estimated through frequency analysis Chen2015 ; Chen2017 .

In this study, the presented estimation process focuses on the dynamic characteristics of a rear-motorized-wheels electric vehicle to achieve the tire road friction estimation and contributes in the following aspect: the estimator is developed by means of a time delay neural network (TDNN) as a way to identify the TRFC based exclusively on the lateral force information. The estimates are compared with a nonlinear least squares (NLS) estimator based on a moving data window.

This paper proceeds as follows: Section 2 presents a vehicle theory development with mathematical models for the tire force models which are used in the estimation method. Section 3 details the least squares regression method. In Section 4, the TDNN estimation algorithm proposed is described. Simulation results are shown and analyzed in Section 5. Finally, this paper is concluded in Section 6.

2 Tire-Ground Contact Model

When sufficient excitations exist in the lateral direction, vehicle lateral dynamics can be used as the basis of the TRFC estimation. The most common tire friction models used in the literature are those of algebraic tire slip angle and force relationships. Although many approaches to the tire-road friction modeling can be found, for this work we selected three analytical models. These models were chosen for their clear and simple formulation. They have fewer tunning parameters and have a good representation of the tire forces nonlinearities.

As mentioned, the force generated between the tire and the road is related to the slip angle and it is of fundamental importance for the knowledge of how the lateral forces arise during a curve. The slip angle

is the angle between the orientation of the tire and the orientation of the velocity vector of the wheel, as depicted in Fig


Figure 1: Schematic diagram of the planar vehicle and tire slip angle representation

The formal definition of slip angle can be derived via kinematic analysis of a planar four-wheel vehicle and is usually defined as:


where and are vehicle longitudinal and lateral velocity components, is the vehicle yaw rate, and are the distances from the vehicle center of gravity to front and rear axles, respectively (as shown Fig. 1) and is half of the wheelbase distance. stand for tire steering angles and the subscripts , , and denote quantities corresponding respectively to the front left, front right, rear left and rear right wheels.

2.1 Mathematical Formulation

Tire models express the relationship between tire forces and moments with slip ratio and slip angle.

Different mathematical tire models have been developed in the literature. The most widely used model is the semi-empirical tire model introduced by Pacejka pacejka2006tyre , called Pacejka tire model or Magic formula. In a simplified form, the formulation of this tire model for lateral force is as follows:


where D, C, B, E are the Magic formula semi-empirical parameters based on tire measurement data, is the vertical offset of the characteristic curve and is the slip angle.

A second model, known as Dugoff’s tire model, was developed in 1969 by Dugoff et al. dugoff1969tire . In its simplest form, the lateral force is expressed as:


where is the normal tire load, is the friction coefficient and the cornering stiffness. Conceptually, cornering stiffness is a property of the tire, experimentally measurable, that changes slowly with time due to tire wear, inflation pressure, and temperature fluctuations hsu2009estimation .

Finally, another widespread model is the Brush model pacejka2006tyre , which defines the lateral force as follows:


Although this article only introduces the most popular and widely used approaches in tire-road friction estimation, there are many valuable studies that have tried to develop new friction models. This subject is addressed in broader texts and books about ground vehicle dynamics such as pacejka2006tyre ; Wong2008theory

The lateral force characteristic curve for each of the presented models is shown in Fig. 2 for several friction coefficients. Initially, the lateral forces increase linearly with the slip angle until it reaches saturation, which represents the tire force limits.

Figure 2: Tire lateral force characteristic curve for each model, varying normal force and three levels of friction

These models show similar behaviors when the slip angle is small. However, they may deviate from each other when high values of vertical force and friction are available. These characteristics suggest that, in the estimation processes, these models could lead to errors due to model discrepancy.

3 Tire-road Friction Identification Through Parametric Regression

As seen in the previous section, the lateral force can be characterized by three fundamental parameters: tire slip angle , normal force and tire-road friction .

When a sufficiently large lateral excitation is detected during a vehicle maneuver, the of road friction estimation can be achieved using the measured signals and the analytical models (2)-(4). This methodology can be seen as a problem of fitting experimental data to a nonlinear analytical function, as addressed in hsu2009estimation ; hsuJournal ; ahn2011robust ; ahnJournal2013 . The method allows the formulation of the problem as one of unconstrained nonlinear least-squares (NLS) optimization.

In other words, we desire to investigate how well we can identify our lateral tire parameters using lateral force information. This requires a good measurement of the lateral tire forces, as well as the knowledge of individual tires. However, if unavailable, the use of an estimate of the axle forces (lumped forces) may hold a lumped friction estimate.

The nonlinear curve-fitting in a least-squares problem consists of finding decision variables that solve the problem:


where is the optimum value that minimizes the objective function, with the parametric function and representing the measured data.

Assuming that the tire analytical models are a good representation of the lateral tire force behavior, they can be used as a parametric function of the NLS method with sets of observed data (in this case, groups of and ).

Despite the promising results of this approach hsuJournal ; ahnJournal2013 , the NLS method has some drawbacks. It requires a long computation time and sometimes this process fails to converge to the true optimal values. The estimator based on NLS generally shows stable estimation results, but does not always guarantee stability and it is difficult to quantify the stability and convergence ahn2011robust . Furthermore, a critical drawback of the NLS is that it is computationally heavy. In a low-speed microprocessor, it may not sustain the same level of performance.

As an alternative to this methodology, we propose a new approach using neural networks. The problem is approached in a similar form, with the same window of observable data applied in a time-delayed neural network.

4 Estimation of Tire-road Friction Coefficient Using Neural Networks

This section proposes a time delay neural network to detect the TRFC. Two main benefits are expected from this method: firstly, a TDNN can establish network connections and the relationship between input and output instead of storing an entire complex tire model in the controller, which can significantly reduce the computations, guarantee the real time performance and avoid model errors due to model discrepancy; secondly, because the TDNN is trained by measured data, it is able to create a mapping from input parameters to the friction coefficient and accurately capture the temporal structure hidden in the data Zhang2017 .

As the analytical models (2),(3) and (4) show, the lateral force is dependent of , and and these are therefore the parameters selected to feed the neural network. Fig. 3 shows the overall structure used for the TRFC estimation.

Figure 3: Block diagram of the proposed hierarchical estimator

Although the existence of tire force sensors, the forces are still hard to be measured and the sensors are very expensive. As solution, a Kalman filter is used for tire force estimation. Here, we use the approach presented in

Cordeiro2016 ; Cordeiro2017 to estimate and using ordinary vehicle sensors, such as GPS, inertial measurement unit (IMU), and encoders. All measures needed for the estimation process are listed in Table 1.

Signal Description
, ,
Vehicle longitudinal, lateral
and vertical accelerations
u, v, w
Vehicle longitudinal, lateral
and vertical linear velocities
p, q, r
Vehicle roll, pitch and yaw
angular rates
Suspension deflection
Tire steering angle
Table 1: Signals and description of measures

With this approach the forces are detected individually, which holds the potential of detecting the TRFC independently for each tire. The wheel slip angle is calculated straightforward using (1).

Also, a supplementary consideration should be taken to ensure the algorithm outputs reasonable estimates. Due to sensors noise and the inherent perturbation on lateral forces, and specially on slip angle calculation, the TDNN inputs should be low-pass filtered to prevent the high frequency disturbances from being propagated to the estimate. A unit gain 5 Hz low-pass filter is applied to the estimated forces and slip angle, as shown in Fig. 3.

Before advancing into the learning process of the proposed neural network, it is important to make one addition to the model. When considering the correlation between friction and each tire measure, the correlation coefficient between the normalized lateral force is significantly higher if compared with each force separately, as listed in Table 2. According to (1), the slip angle is determined by vehicle velocities and normal force is mainly affected by the roll over effect. These measures are only affected indirectly by friction, thereby, a low correlation is expected.

1.681e 0.2955 1.186e 0.4076
Table 2: Correlation coefficients between the specified variables

Given this fact, the normalized forces should be selected as one input to feed the neural network instead of and separately. The basis for this choice also lies on the friction circle concept in which the maximum value of the resultant force is determined along a circle (directly influenced by friction), and this value can be decomposed into the limits of the normalized forces Gim1991 .

It is important to note that a positive correlation was observed for positive lateral forces (data obtained from a right hand maneuver). A negative correlation with similar magnitude is expected for negatives forces. The signal, thus, is a consequence of the reference frame.

The TDNN architecture is depicted in Fig. 4. Two inputs were selected: a normalized lateral force , obtained from the kalman estimator, and the calculated slip angle.

Figure 4: Time-delayed neural network architecture with a single hidden layer

The configuration of the proposed TDNN for TRFC is as follows (see Fig. 4): 2 inputs with 50 samplings delay (observation window of size N

= 50) and one single hidden layer with 50 neurons. The neurons differentiable transfer function is nonlinear, properly selected as a Tan-Sigmoid transfer function.

In the neural network data collecting stage, about 200,000 original data are obtained from simulation with a 100 Hz sampling rate (translating to an observation window of 0.5 seconds). The range of variation of the network input parameters to the tire model is bounded as described in Table 3. The friction coefficient is set with different levels and the vehicle response (data of , and ) is obtained.

Input parameter     Variation
Friction coefficient      0.3 to 1.2 at intervals of 0.1
Slip angle      [-0.12 0.12] rad
Lateral Force      [-2.8 2.8] kN
Normal force      [2 4.4] kN
Table 3: Data training parameters and space dimension

As in the classical Neural Network, the Time Delay Neural Network also has a training phase. Training was carried out using Matlab Neural Network Toolbox. The Levenberg-Marquardt

algorithm is used with 1000 epochs of training iterations with 70% of the collected data randomly taken as the training set, 15% used for validation set and 15% as the test set.

5 Results

The simulation results presented in this section are obtained using Matlab/Simulink. A representative and realistic full-vehicle multibody dynamics model (including a steering system, powertrain system, suspension system and the Pacejka tire model for tire ground interactions), was used consisting of the following motions:

  • Longitudinal, lateral and vertical body motion;

  • Wheels rotation;

  • Unsprung masses motion;

  • Pitch, roll and yaw body rotation.

The physical parameters of the car used in this study (validated and extracted from rafaelcordeirothesis ) are listed in Table 4. The vehicle model is used to simulate a real rear wheel drive vehicle, providing references of a vehicle state and measured signals. Gaussian noises are added (according to the commercial MTi Xsens sensor specifications (MTi-G-700)) in the simulated measurements to realistically reproduce a real application.

Parameter name     Value
Vehicle mass      1100 kg
Yaw inertia moment      1350
Roll inertia moment      337.5
Pitch inertia moment      1350
Distance from CG to front wheels      1.5 m
Distance from CG to rear wheels      1.9 m
Wheelbase      1.8 m
Wheel rotational inertia      1
Wheel radius      0.25 m
Height of CG      0.5 m
Table 4: List of vehicle main physical parameters

The simulation results of three representative maneuvers are presented here. Table 5 gives the details and purpose of each maneuver. Fig. 5 provides the physical representation of each proposed scenario where the color designates the change in friction.

Each case was performed on a different theoretical surface, where a theoretical = 1.0 surface roughly corresponds to driving on a dry pavement, = 0.8 on a wet pavement and = 0.6 corresponds to driving on gravel Wang2004 .

Simulated Maneuver Test Surface Purpose
Ramp steer with constant
friction coefficient
(Fig. (a)a)
Constant dry pavement
( = 1)
Investigate the early friction sensing ability of the
estimator. During the maneuver, lateral forces are
gradually increasing, eventually reaching saturation.
Constant steer - case 1:
step varying for all tires
(Fig. (b)b)
Five levels of TRFC
with unequal steep sizes
( = 1.0 to 0.6)
Determine the estimator response to a quick
change in road surface. Friction coefficient is
varied at five new levels.
Constant steer - case 2:
different step varying
for left/right tires
(Fig. (c)c)
Each wheel is subjected
simultaneously to two
different frictions
( = 0.9, 0.8, = 0.8, 0.7)
Validate the TRFC detection under small friction
variations. Also, each wheel is exposed to different
surfaces at the same time (distinct for
FL and FR tires).
Table 5: Details and investigative purpose of each simulated maneuver
Figure 5: Schematic road layout with changing in friction coefficient

5.1 Ramp Steer Maneuver

Fig. 6 displays the data resulting from the simulation of a left-hand ramp steer maneuver. The steering angle goes linearly from 0 to -18 degrees at the roadwheels reference.

Figure 6: Front right lateral force and slip angle for a ramp steer maneuver

To ensure that there is enough data to be meaningful for the NLS fit and the TDNN approach, first the process is initialized by placing a slip angle threshold . The slip angle data must exceed before the estimation begins, otherwise the fitting optimization may not guarantee a reliable solution.

The TDNN estimator will be here compared with the NLS approach. To show the dependency of the NLS fit with the mathematical model, the regression is performed choosing the Dugoff and Brush models as a parametric function of the nonlinear regression. The window is selected with size and will be used in all cases showed from here on.

Using the estimation algorithm with = 1 degree, the NLS and TDNN algorithm waits until the front tire slip angle exceeds at t = 12 s before fitting the force-slip data (see Fig. 7). The estimated value is the optimum solution of the optimization problem (6) and before instant 12 s the estimation simply holds at initial value .

Figure 7: Friction estimates from (,) data for different approaches

The slight increase in the friction estimate as the maneuver progresses is expected. Initially, during linear tire regime operation, lateral forces measurements have yet to reach their peak value and both methods underestimate the friction coefficient. As more lateral force measurements become available, the peak force limit is reached and the friction estimate reaches a final estimate. Therefore, adequately large slip angles are required for stable and accurate estimation in both TDNN and NLS methods and the slip angle data threshold is indispensable.

The model error also becomes apparent on the NLS fit, where Brush and Dugoff models show different convergence values due to model discrepancy. This divergence arises due to the difference between Pacejka model, used to generate the data, with the models used on the estimators (as discussed in section 2).

5.2 Constant Steer: Case 1

In this experiment, the tire-road friction coefficient is set at five levels, varying randomly from 1.0 to 0.6. The transitions occur during successive equal time intervals of 10 seconds. The vehicle is set on an equilibrium point in a constant left turn maneuver. The steering angle of the front left and right tires is set to -18.36 and -15.82 degrees. These values follow the steering Ackerman Geometry. As a consequence, the inner-turn wheel reaches higher tire side slip and lateral force if compared to the outer-turn wheel, as shown in Fig. 8. Also, the vertical force is higher on the right tire. The load transfer appears as a result of the roll over effect and is kept constant due to the static steer maneuver nature.

(a) front left (fl)
(b) front right (fr)
Figure 8: Slip angle, vertical and lateral forces of the tires: Constant steer case 1

The TRFC estimation results are shown in Fig. 9. One can note the front left estimative is more accurate and less oscillating than the front right estimative. A necessary condition for good estimation results, as shown previously, is a large lateral excitation (high slip angle).

Figure 9: Front left and front right friction estimates in a time varying friction scenario

This experiment highlights the dependency of the NLS to the parametric function. The NLS final estimates of each interval show a constant error bias while the TDNN produces a solid and concise estimative. Naturally, there are discrepancies between the dynamic behavior of the real tire system and the derived mathematical model (see Fig. 2). Therefore, a constant error bias should be expected on the NLS model-based approach.

Table 6 displays the root mean square (RMS) error of the estimates of the front left and right tires. Although very similar, a high overall estimation accuracy is achieved for both techniques and it shows great promise for a real implementation.

TDNN 0.0346 0.0350
0.0421 0.0357
0.0546 0.0395
Table 6: RMS error of the estimated friction - case 1

Moreover, the rate of convergence is slightly higher on the TDNN approach. Figure 10 highlights this response by zooming in Figure (a)a on three intervals of friction transitions. The TDNN estimates converges to a more accurate values with faster responses than the NLS estimates. This behavior lies in the fact that the relationship between input and output was correctly mapped on the database and therefore can be observed on the following results.

Figure 10: Friction estimates rate of convergence comparison

5.3 Constant Steer: Case 2

While case 1 showed an equal change for both wheels, here, the tire-road friction coefficient is set at two different levels for each tire. The transition occur during an equal time interval of 10 seconds. Left wheel friction undergoes a transition from 0.9 to 0.8 while right TRFC goes from 0.8 to 0.7.

The vehicle is set on an equilibrium point in a constant left turn maneuver. The knowledge of forces and friction of individual tires is desirable and would offer stability control systems with most needed information. Thus, this change maneuver is conducted to verify that the estimator can identify the friction for each tire individually and certify that the estimation is indeed independent for each wheel.

Fig. 11 displays the simulated slip angle, lateral and vertical forces from the proposed right-hand steer maneuver. At instant 10 s, the road surface adhesion coefficient decreases to a different value for each tire. Again, vertical forces are maintained constant due to the maneuver nature, with the changing in friction mostly affecting lateral forces and slip angle.

(a) front left (fl)
(b) front right (fr)
Figure 11: Slip angle, vertical and lateral forces of the tires: Constant steer case 2

Fig. 12 shows the estimation results. With the TDNN approach, the individual wheel friction is confidently estimated with high accuracy. Since slip angles are larger than for the case 1 scenario, it satisfies the required large lateral forces excitation and gives very accurate estimates. On the NLS estimation results, however, a constant error are still apparent and can be seen during the transitions. This characteristic should be considered when the estimated result is used for control purposes.

Figure 12: Front left and front right friction estimates in a time varying friction scenario with different friction for each wheel

The TDNN shows a slightly better estimation quality, as seen in the RMS error listed in Table 7. Note that the regression based method still exhibits the hindsight bias caused by the inevitable model differences.

TDNN 0.0638 0.0477
0.0652 0.0493
0.0666 0.0486
Table 7: RMS error of the estimated friction - case 2

6 Conclusions

In this study, we presented a hierarchical TRFC estimation method based on a time delay neural network and compared it with a classical nonlinear regression approach, using the same data observation window. The overall estimation algorithm was evaluated on varying road surfaces with three different scenarios using the Matlab/Simulink platform.

Although road-friction was accurately identified using both algorithms, there is a primary shortcoming in the presented lateral-force based friction estimation: it requires sufficient levels of lateral excitation for the correct friction identification. An earlier knowledge of the TRFC is desirable, however, both approaches showed a similar behavior: a necessary waiting time for the tire slip angle to fulfill the observation window satisfying the specified excitation threshold.

Nonetheless, as algebraic methods, the NLS method rely more heavily on accurate models and is a major reason for estimation errors. The NLS need a very reliable and trustworthy parametric function with precise tunning parameters and stationary estimation errors should be expected.

On the other hand, the TDNN method is independent of any mathematical tire model, however, requires a sufficient and representative database. In this study, the TDNN was also able to provide estimates with lower RMS errors compared with the classical regression approach. It also demands less computation time at each time instant and may be the best alternative for a real time implementation in embedded systems.

Since the proposed method is only analyzed theoretically and validated via simulation, an actual benchmark or field test is needed in the subsequent work to verify the proposed approach. Future works may also include the design of a neural network containing not only lateral information but also longitudinal forces, slip ratio and self-moment align.

This work was supported by FCT Portugal, through IDMEC under projects LAETA (UID/EMS/ 50022/2019). The mobility of A. Ribeiro has been possible with the Erasmus Mundus SMART support (Project Reference: 552042-EM-1-2014-1-FR-ERA MUNDUS-EMA2) coordinated by CENTRALESUPELEC. The authors also acknowledge the support of FAPESP through Regular project AutoVERDE N. 2018/04905-1, Ph.D. FAPESP 2018/05712-2 and CNPq grant 305600/2017-6.


  • (1) Ahn, C., Peng, H., Tseng, H.: Robust estimation of road frictional coefficient. IEEE Transactions on Control Systems Technology 21(1), 1–13 (2013). DOI 10.1109/TCST.2011.2170838
  • (2) Ahn, C.S.: Robust estimation of road friction coefficient for vehicle active safety systems. Ph.D. thesis, University of Michigan (2011)
  • (3) Al-Zughaibi, A.I.H.: Experimental and analytical investigations of friction at lubricant bearings in passive suspension systems. Nonlinear Dynamics 94(2), 1227–1242 (2018). DOI 10.1007/s11071-018-4420-x. URL
  • (4) Casselgren, J., Sjodahl, M., Woxneryd, S., Sanfridsson, M.: Classification of road conditions – to improve safety. In: 11th International Forum on Advanced Microsystems for Automotive Applications, pp. 47–59. Springer Science Business Media (2007)
  • (5) Chen, L., Bian, M., Luo, Y., Qin, Z., Li, K.: Tire–road friction coefficient estimation based on the resonance frequency of in-wheel motor drive system. Vehicle System Dynamics 54(1) (2015)
  • (6) Chen, L., Luo, Y., Bian, M., Qin, Z., Luo, J., Li, K.: Estimation of tire-road friction coefficient based on frequency domain data fusion. Mechanical Systems and Signal Processing 85, 177–192 (2016)
  • (7) Cordeiro, R.A., Ribeiro, A.M., Oliveira, J.R., Victorino, A.C., Ferreira, P.A., de Paiva, E.C., Bueno, S.S.: Road grades and tire forces estimation using two-stage extended kalman filter in a delayed interconnected cascade structure. IEEE Intelligent Vehicles Symposium (2017)
  • (8) Cordeiro, R.A., Victorino, A.C., Ferreira, P.A., de Paiva, E.C., Bueno, S.S.: Tire-ground forces estimation in a 4-wheel vehicle using a delayed interconnected cascade-observer structure. IFAC-PapersOnLine 49(15), 139–144 (2016)
  • (9) Cordeiro, R.d.A.: Track control of a four-wheel robotic vehicle in adverse terrain conditions. Ph.D. thesis, Universidade Estadual de Campinas (2017)
  • (10) de Castro, R., Araujo, R.E., Cardoso, J.S., Freitas, D.: A new linear parametrization for peak friction coefficient estimation in real time. In: 2010 IEEE Vehicle Power and Propulsion Conference, pp. 1–6 (2010). DOI 10.1109/VPPC.2010.5729138
  • (11) Dugoff, H., Fancher, P.S., Segel, L.: Tire performance characteristics affecting vehicle response to steering and braking control inputs. Tech. Rep. Contract CST-460, Office of Vehicle Systems Research (1969)
  • (12) Gim, G., Nikravesh, P.E.: An analytical model of pneumatic tyres for vehicle dynamic simulations. part 2: Comprehensive slips. International Journal of Vehicle Design 12(1), 19–39 (1991)
  • (13) Hsiao, T., Liu, N.C., Chen, S.Y.: Robust estimation of the friction forces generated by each tire of a vehicle. In: Proceedings of the 2011 American Control Conference, pp. 5261–5266 (2011). DOI 10.1109/ACC.2011.5991219
  • (14) Hsu, Y.H.J.: Estimation and control of lateral tire forces using steering torque. Ph.D. thesis, Stanford University (2009)
  • (15) Hsu, Y.H.J., Laws, S., Gadda, C.D., Gerdes, J.C.: A method to estimate the friction coefficient and tire slip angle using steering torque. In: ASME 2006 International Mechanical Engineering Congress and Exposition, pp. 515–524. American Society of Mechanical Engineers (2006)
  • (16) Jokela, M., Kutila, M., Le, L.: Road condition monitoring system based on a stereo camera. In: 2009 IEEE 5th International Conference on Intelligent Computer Communication and Processing, pp. 423–428 (2009). DOI 10.1109/ICCP.2009.5284724
  • (17) Jonsson, P.: Classification of road conditions: From camera images and weather data. In: International Conference on Computational Intelligence for Measurement Systems and Applications (CIMSA) Proceedings. Institute of Electrical and Electronics Engineers (IEEE) (2011)
  • (18) Judy Hsu, Y.H., M. Laws, S., Christian Gerdes, J.: Estimation of tire slip angle and friction limits using steering torque 18, 896 – 907 (2010)
  • (19) Li, K., Misener, J.A., Hedrick, K.: On-board road condition monitoring system using slip-based tyre-road friction estimation and wheel speed signal analysis. Proceedings of the Institution of Mechanical Engineers, Part K: Journal of Multi-body Dynamics 221(1), 129–146 (2007). DOI 10.1243/1464419JMBD60
  • (20) Müller, S., Uchanski, M., Hedrick, K.: Estimation of the maximum tire-road friction coefficient. Journal of Dynamic Systems, Measurement, and Control 125(4), 607 (2003). DOI 10.1115/1.1636773
  • (21) Pacejka, H.B.: Tyre and vehicle dynamics, second edn. Butterworth-Heinemann, Oxford (2006)
  • (22) Pasterkamp, W.R., Pacejka, H.B.: The tyre as a sensor to estimate friction. Vehicle System Dynamics 27(5-6), 409–422 (1997)
  • (23) Qi, Z., Taheri, S., Wang, B., Yu, H.: Estimation of the tyre–road maximum friction coefficient and slip slope based on a novel tyre model. Vehicle System Dynamics 53(4), 506–525 (2015). DOI 10.1080/00423114.2014.1002795
  • (24) Rajamani, R., Phanomchoeng, G., Piyabongkarn, D., Lew, J.Y.: Algorithms for real-time estimation of individual wheel tire-road friction coefficients. IEEE/ASME Transactions on Mechatronics 17(6), 1183–1195 (2012). DOI 10.1109/TMECH.2011.2159240
  • (25) Shao, L., Jin, C., Lex, C., Eichberger, A.: Robust road friction estimation during vehicle steering. Vehicle System Dynamics 57(4), 493–519 (2019). DOI 10.1080/00423114.2018.1475678. URL
  • (26) Singh, K.B., Taheri, S.: Estimation of tire–road friction coefficient and its application in chassis control systems. Systems Science & Control Engineering 3(1), 39–61 (2014)
  • (27) Tuononen, A.: Optical position detection to measure tyre carcass deflections. Vehicle System Dynamics 46(6), 471–481 (2008). DOI 10.1080/00423110701485043
  • (28) Wang, J., Alexander, L., Rajamani, R.: Friction estimation on highway vehicles using longitudinal measurements. Journal of Dynamic Systems, Measurement, and Control 126(2), 265 (2004). DOI 10.1115/1.1766028
  • (29) Wong, J.Y.: Theory of Ground Vehicles, third edn. Wiley, Hoboken, N.J (2008)
  • (30) Xia, X., Xiong, L., Sun, K., Yu, Z.P.: Estimation of maximum road friction coefficient based on lyapunov method. International Journal of Automotive Technology 17(6), 991–1002 (2016). DOI 10.1007/s12239-016-0097-7
  • (31) Zhang, X., Göhlich, D.: A hierarchical estimator development for estimation of tire-road friction coefficient. PLOS ONE 12(2), e0171,085 (2017)