686 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 47, NO.

5, OCTOBER 2017

Nonlinear Driver Parameter Estimation and Driver

Steering Behavior Analysis for ADAS Using
Field Test Data
Changxi You, Jianbo Lu, and Panagiotis Tsiotras, Senior Member, IEEE

Abstract—In the development of advanced driver-assist systems automotive active safety systems, these studies show that driver
(ADAS) for lane-keeping or cornering, one important design objec- behavior still remains the most important factor contributing to
tive is to appropriately share the steering control with the driver. accidents. It is therefore necessary to understand, characterize
The steering behavior of the driver must therefore be well char-
acterized for the design of a high-performance ADAS controller. and, if possible, predict driver behavior so as to design better,
This paper adopts the well-known two-point visual driver model and more proactive (as opposed to merely reactive) advanced
to characterize the steering behavior of the driver, and conducts driver-assist systems (ADAS). Nevertheless, driver modeling is
a series of field tests to identify the model parameters and vali- a difficult task since driver behavior is affected by different in-
date this model in real-world scenarios. An extended Kalman filter dividual factors, such as gender, age, experience, and driver’s
and an unscented Kalman filter are implemented for estimating
the driver parameters using either a joint-state estimation algo- aggression. Such diverse driver behaviors have a significant ef-
rithm or a dual estimation algorithm. The estimated parameters fect on the performance of ADAS [4], [5]. A controller for
for different types of drivers are analyzed and compared. The re- vehicle handling stability should take into account the diverse
sults show that the two-point visual driver model captures realistic driver skills, habits, and handling behavior of different drivers,
driving behavior with time-varying, but not necessarily constant, and persistently provide good “intuitive” performance. In order
parameters. A wavelet analysis of the driver steering command
shows that distinct driver classes can be identified by analyzing the to characterize driver behavior, researchers have proposed dif-
smoothness of the driver command using the Lipschitz exponents ferent driver models based on several methodologies over the
of the recorded signals. past four decades.
Index Terms—Extended Kalman filter (EKF), field test,
Wier and McRuer [6] used transfer functions to describe the
parameter estimation, two-point visual driver model, unscented result of the driver’s actions on the vehicle’s position error and
Kalman filter (UKF), wavelet signal analysis. yaw angle, and built a quasi-linear model (crossover model) to
approximately describe the nonlinear steering behavior of the
I. INTRODUCTION driver. This model uses feedback control to eliminate the track-
ORE than six million motor vehicle crashes occurred in ing error, but it does not take the driver’s preview behavior into
M the U.S. in 2014 alone, of which 27% resulted in injury
or death [1]. From 2014 to 2015, the total number of vehicle
consideration. MacAdam [7], [8] assumed that the driver wants
to minimize a predefined previewed output error, and modeled
crashes increased by 3.8%, and the number of fatal crashes in- the driver’s steering strategy as an optimal preview process with
creased by 7% [2]. Another study, sponsored by National High- a time lag. Hess and Modjtahedzadeh [9], [10] introduced a
way Traffic Safety Administration, investigated 723 crashes and control-theoretic model for the steering behavior of the driver.
showed that driver behavioral error caused or contributed to This model consisted of a preview component along with low-
99% of these crashes [3]. Given the increased sophistication of and high-frequency compensation elements. The above models
successfully achieve lane-tracking using only lateral control;
braking is not considered in these works. Burgett and Miller
Manuscript received June 17, 2016; revised February 23, 2017 and April [11] designed and optimized a parameterized driver model us-
21, 2017; accepted June 10, 2017. Date of publication July 25, 2017; date of
current version September 15, 2017. This work was supported by National ing a multivariable nonlinear regression approach, based on data
Science Foundation Awards CMMI-1234286 and CPS-1544814 and the Ford collected from test tracks and driving simulations. This model
Motor Company. This paper was recommended by Associate Editor Dr. Rafael investigated the driver’s braking strategy in order to avoid rear-
Toledo. (Corresponding author: Panagiotis Tsiotras)
C. You is with the School of Aerospace Engineering, Georgia Institute of end driving conflicts. Chatzikomis and Spentzas [12] proposed
Technology, Atlanta, GA 30332-0150 USA (e-mail: [email protected]). a path-following driver model that regulated both the steering
J. Lu is with the Research and Advanced Engineering, Ford Motor Company, wheel and the throttle/brake by previewing the path ahead of the
Dearborn, MI 48121 USA (e-mail: [email protected]).
P. Tsiotras is with the School of Aerospace Engineering and the Institute for vehicle. In [13] and [14], model predictive controller (MPC)-
Robotics and Intelligent Machines, Georgia Institute of Technology, Atlanta, based driver steering models have been considered. Keen and
GA 30332-0150 USA (e-mail: [email protected]). Cole [14], in particular, linearized the vehicle model at different
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. working points and used a multimodel structure to characterize
Digital Object Identifier 10.1109/THMS.2017.2727547 the ability of the driver to predict the future vehicle path. By

2168-2291 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
YOU et al.: NONLINEAR DRIVER PARAMETER ESTIMATION AND DRIVER STEERING BEHAVIOR ANALYSIS 687

using different combinations of the internal models, this MPC The two-point visual driver model used in this paper is de-
achieves various driver expertise in the path-following task. rived from the concept of the two-level steering mechanism
The driver’s mental work has also been taken into consid- observed in a series of psychological experiments involving
eration for driver modeling. In [15], Flad et al. proposed a human drivers [25]–[27]. In [25], Donges divided the driver’s
steering-primitive optimal selection driver model by defining steering task into a guidance level and a stabilization level, and
a set of elementary control primitives to describe the driver’s thereby built a two-level steering model. The guidance level
neuromuscular system, limbs, and control actions. This model interprets the driver’s perceptual response with respect to the
assumes that the driver has a mental model of the vehicle and oncoming road in an anticipatory open-loop control mode. The
the steering task and determines the optimal sequence of control stabilization level interprets the driver’s compensatory behav-
primitives to achieve the target maneuver. Different artificial in- ior with respect to the deviation from the reference path in a
telligence approaches have also been introduced to model the closed-loop control mode. This idea has been widely accepted
driver’s mental work and behaviors. In [16], Kageyama and and has been further developed by subsequent researchers [22],
Pacejka evaluated the driver’s mental influence from the envi- [23], [26]–[28]. Among these researchers, Salvucci [23] first in-
ronment with respect to a “risk level” and proposed a driver troduced the concepts of visual “near point” and “far point” into
model based on fuzzy control theory. Lin et al. [17] built a the model. By taking appropriate choices of the “near point” and
neural network driver model and compared three typical model “far point,” the two-point visual driver model achieves different
configurations in great detail. More recently, Hamada et al. [18] tasks such as lane tracking [23] and collision avoidance [29].
proposed a beta process autoregressive hidden Markov model The contributions of this paper can be summarized as fol-
(HMM). This model was trained in an unsupervised way using lows: First, the paper adopts the two-point visual driver model
real driving data, and was used to predict the driving behaviors from [22], since this model characterizes driver steering behav-
of the drivers. ior more precisely. This driver model combines both a two-level
All previous driver control-theoretic models can be catego- visual strategy and high-frequency kinesthetic feedback. The
rized into the following three groups according to the method- latter accounts for the interaction between the driver’s arms and
ology used to develop them. the steering wheel [9]. Saleh et al. in [21], [30], [31] also adopted
1) Classical control theory such as [6], [9], and [10], where the two-level visual strategy, but instead of the high-frequency
the system is represented using transfer functions and the kinesthetic feedback in [9], [22], a well-designed neuromuscu-
stability is analyzed using frequency-response methods. lar system was used. The identification of the parameters of the
2) Modern control theory such as [7], [8], [11], [12], and model in [30], [31], and [21] was done using simulated data. In
[14], where the system is represented in state space and this paper, we show the validity of the proposed model by com-
the stability is analyzed in the time domain. paring with actual recorded driver data collected during field ex-
3) Intelligent control theory such as [16]–[18], where the ar- periments. Although previous work has validated the two-point
tificial intelligence approaches including neural network, visual driver model and identified the driver model parameters
fuzzy logic, and HMM are used to develop the driver using a driving simulator [22], [30], this is the first instance that
models [19]. the model is validated using actual field test data. Second, by
These driver models focus on three kinds of driving tasks, in- applying four different identification methods, namely, the joint
cluding longitudinal control [11], lateral control [6]–[10], [14], extended Kalman filter/unscented Kalman filter (EKF/UKF) and
[15], [17], and combined longitudinal-lateral control [12], [16], the dual EKF/UKF [32]–[34] it is shown that the model param-
[18]. eters are indeed identifiable using minimal data, but that some
Recently, nonparameterized models such as neural networks of these parameters are not necessarily constant but may vary
or HMMs have been used to predict driver behavior. They have with time. Our results thus reveal that parameter-varying ver-
to be trained offline by using supervised/unsupervised machine sions of the two-point visual driver model may provide a much
learning techniques and they typically need large amounts of better explanation of actual human driver behavior. It is expected
data. Furthermore, nonparameterized models are not very trans- that these observations will pave the way for online driver be-
parent to the user and hence are not convenient for design- havior and cognitive driver state identification, which can be
ing driver-based ADAS controllers. The parameters of these used downstream in the ADAS architecture in order to adapt
models are difficult to modify in order to characterize different the controller gains to the specific driver/vehicle/traffic configu-
driving behaviors; instead, the model must be retrained using ration. Finally, we show that when comparing different driving
new data to capture new driver types and driving styles. Pa- types, the smoothness of the driver steering command may be
rameterized, transfer-function-based driver models, such as the a good discriminating feature for driver classification. Using
crossover model [6], [20], the control theoretic model [9], and wavelet signal analysis, it is shown that different driver styles
the two-point visual driver model [21]–[23] on the other hand correspond to different signal smoothness (i.e., degree of dif-
are better for control design tasks, since they are easy to use (they ferentiability), as measured by the rate of decay of the wavelet
are quasi/linear), and their parameters correspond to measurable coefficients. As far as we know, this is the first work that wavelet
physical variables that relate to meaningful performance param- analysis has been applied to determine driver categories.
eters. Among these driver models, the two-point visual driver The paper is structured as follows. Section II introduces the
model is considered to have both satisfactory model accuracy mathematical modeling of the driver. Section III details the ap-
and good identification feasibility [24]. proaches used to identify the driver model parameters, while

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
688 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 47, NO. 5, OCTOBER 2017

Fig. 1. Human-vehicle-road closed-loop system.

Section IV describes the equipment and the driving scenarios

used for the field tests. Section V outlines the data processing
task and presents the results. Section VI analyzes and compares
different driver styles. Finally, Section VII summarizes the re-
sults of this study and provides some directions for future work.
Fig. 2. Road geometries, vehicle states, and driver’s visual perception.
II. SYSTEM MODELING AND PROBLEM FORMULATION
The proposed human-vehicle-road system consists of four functions of the blocks shown in Fig. 1 are given below
subsystems, as shown in Fig. 1.
TL s + 1
1) The driver model that exerts a steering torque on the steer- Ga (s) = Ka , Gc (s) = Kc
ing wheel. TI s + 1
The steering column model that converts steering torque 1 Tk1 s
Gnm (s) = , Gk1 (s) = KD
to steering angle; TN s + 1 Tk1 s + 1
The vehicle model that provides the necessary position Tk2 s + 1
and state information of the vehicle; and GL (s) = e−t p s , Gk2 (s) = KG (1)
Tk3 s + 1
The road and perception model that provides the road
geometry and kinematics, and also determines the driver’s where Ka and Kc are static gains for the anticipatory and com-
visual perception angles. pensatory control subsystems, respectively; KD and KG are
The input to the system is the curvature of the road ρref , static gains for the kinesthetic perception feedback subsystems,
which can be treated either as an external reference command respectively; TL and TI (TL > TI ) are the lead time and lag time
to be tracked or a disturbance to be rejected, depending on the constants, respectively; Tk1 , Tk2 , and Tk3 are the three time con-
problem formulation. The primary performance variable is the stants of the driver’s kinesthetic perception feedback from the
lateral deviation Δy of the so-called “near point” directly in steering wheel, tp is the delay for the driver to process sensory
front of the vehicle to the centerline of the road (see Figs. 1 and signals, and TN is the time constant of the driver’s arm neuro-
2). muscular system. Ka , Kc , KD , KG , TL , TI , TN , Tk1 , Tk2 , Tk3 ,
and tp are the 11 parameters of the driver model.
A. Driver Model
B. Road and Perception Model
We use the driver model proposed in [22], which introduces
a kinesthetic force feedback from the steering wheel. The struc- The road and perception model interacts with both the vehicle
ture of this model is shown in the red rectangular box in Fig. 1. model and the driver model (refer to Fig. 1) and achieves two
The transfer functions Ga (s) and Gc (s) account for the anticipa- functions: 1) It determines the vehicle’s position and posture
tory control and the compensatory control actions of the driver, relative to the road geometry; and 2) it determines the loca-
respectively. The system Gnm (s) approximately describes the tion of the driver’s near and far visual points on the upcoming
neuromuscular response of the driver’s arms. The “Delay” road. The near visual point is fixed at a certain distance along
block indicates the driver’s processing delay in the brain, and the heading direction of the vehicle, while the far visual point
the transfer functions Gk1 (s) and Gk2 (s) account for the driver’s is taken as the tangent point on the inner road boundary for
kinesthetic perception of the steering system. The variables Tant driving on a curved road, or the vanishing point of the road for
and Tcom denote the driver’s steering torques corresponding to driving along a straight road [23]. Fig. 2 illustrates the relations
the anticipatory control and the compensatory control paths, re- between the geometry of driver’s visual perception, the vehicle
spectively; δs denotes the steering wheel angle; and the inputs and the curved road [35], [36]. In Fig. 2, the frame XI -O-YI
θnear and θfar denote the near-field and the far-field visual an- is fixed on the road. It is assumed that the vehicle is cornering
gles, respectively (see Fig. 2). Finally, Tdr denotes the driver’s with a certain lateral deviation from the road centerline. Let
total steering torque delivered at the steering wheel. The transfer ψ denote the vehicle’s yaw angle, let ψt denote the angle be-

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
YOU et al.: NONLINEAR DRIVER PARAMETER ESTIMATION AND DRIVER STEERING BEHAVIOR ANALYSIS 689

tween the tangent to the road centerline and the XI axis, and let parameter is restricted to lie within some compact interval, pi ∈
M denote the current position of the vehicle’s center of mass. [pi , pi ], i = 1, 2, . . . , 12. Let p = (p1 , p2 , . . . , p12 )T ∈ P =
Let also A denote the driver’s “lookahead” point in front of [p1 , p1 ] × [p2 , p2 ] × · · · × [p12 , p12 ] ⊂ R12 . The upper and
M at a distance s along the vehicle’s heading direction, let B lower bounds (pi and pi ) that define P are given in Table III.
denote the intersection of OA with the road centerline, let E The combined system of the driver model and the road and
denote the intersection of AB with the tangent to the road cen- perception model can be written in the form
terline, and let C denote the point of tangency of the line along
the gaze direction on the road’s inner boundary. Furthermore,
ẋc = Ac (p)xc + B c (p)uc (3a)
let Ls denote the distance between C and M, let θfar denote the
visual angle between the gaze direction of the driver from a far y =C x
c c c
(3b)
away point and the heading direction of the vehicle, and let θnear
denote the near-point visual angle between MB and the heading where the system state is xc = (Δψ, δy, xd1 , xd2 , Tdrff ,
direction of the vehicle. Finally in Fig. 2, Δy denotes the length xd3 , xd4 , Tdrfb )T , the input is uc = (ρ, β, r, δs )T , and the output
of the line segment AB—the predicted deviation from the road is y c = Tdrff + Tdrfb = Tdr . In the previous expressions, Tdrff and
centerline at the near lookahead point if the vehicles continues Tdrfb denote the two components of the driver’s steering torque,
with the current heading, Rref denotes the radius of the road’s resulting from the feedforward path and the feedback path of
inner boundary, d denotes the distance from M to the road’s in- the driver model, respectively. Specially, referring to Fig. 1, Tdrff
ner boundary, and D denotes the width of the road. Henceforth, and Tdrfb can be expressed as follows:
it will be assumed that d and D are small compared to Rref .
From Fig. 2, the near- and far-distance visual perception angles
can be approximated as [22], [25], [35]–[38] Tdrff = (Tcom + Tant )GL Gnm (4a)

Δy Tdrfb = −δs Gk1 (1 + Gk2 )Gnm . (4b)

θnear ≈ (2a)
s
Ls By measuring uc and y c , we can identify the driver parameter
θfar ≈ + Δψ ≈ Ls ρref + Δψ (2b) vector p in (3a) and (3b). To this end, we define an alternative
Rref
parameter vector ν = (ν1 , ν1 , . . . , ν12 )T as follows:
where ρref = 1/Rref is the road curvature, and Δψ = ψt − ψ is
the angle between the tangent of the road centerline and the
1 1 1 p1
vehicle’s heading direction. ν1 = , ν2 = , ν3 = , ν4 =
p4 p6 p5 p5
C. Problem Formulation p2 p3 p2
ν5 = , ν6 = , ν7 = p7 , ν8 = p8
p4 p6 p7 p4 p7
We formulate the driver parameter estimation problem based
on the driver model, road and perception model, steering 1 1
ν9 = p9 , ν10 = , ν11 = p11 , ν12 = . (5)
column model, and vehicle model summarized in the pre- p10 p12
vious section. For notational simplicity, let p1 = Ka , p2 =
Kc , p3 = TL , p4 = TI , p5 = TN , and p6 = tp . The driver’s The mathematical expressions in the sequel can be simplified
near-field lookahead distance s is also an important fea- by using ν instead of p. The system matrices in (3a) and (3b) are
ture of the driver steering characteristics. We thus take s given explicitly by (6). It is worth mentioning that, Vx is assumed
as an additional parameter, and let p7 = s . We further let to be constant in (6). One can add Vx to the input vector uc for
p8 = KD , p9 = KG , p10 = Tk1 , p11 = Tk2 , and p12 = Tk3 for varying velocity cases.
the high-frequency kinesthetic feedback in the driver model. Since we are interested in identifying the parameter vector ν,
Since the human driver has physical limits, each model we augment the state with ν and define the new augmented state

 
Ac (ν) | B c (ν)
=
Cc | 0
⎡ ⎤
0 0 0 0 0 00 0 | Vx 0 −1 0
⎢ Vx 0 0 0 0 00 0 | Vx ν7 −Vx −ν7 0 ⎥
⎢ ⎥
⎢ 0 ν6 − ν ν1 ν 5 −ν1 0 0 00 0 | 0 0 0 0 ⎥
⎢ ν ν 2 ⎥
⎢4 2 4 4ν 4ν −2ν 0 0 0 0 | 4 L s νν 32 ν 4 0 0 0 ⎥
⎢ ν3 5 2 2 ⎥
⎢ −ν4 − ν 5 ν 3 −ν3 ν3 −ν3 0 0 0 | −Ls ν4 0 0 0 ⎥. (6)
⎢ ν2 ⎥
⎢ 0 0 0 0 0 00 −ν3 ν10 ν12 | 0 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 0 1 0 −ν3 ν10 − ν3 ν12 − ν10 ν12 | 0 0 0 −ν3 ν8 ν12 (ν9 + 1) ⎥
⎢ ⎥
⎣ 0 0 0 0 0 01 −ν3 − ν10 − ν12 | 0 0 0 −ν3 ν8 (ν9 ν11 ν12 + 1) ⎦
0 0 0 0 1 00 1 | 0 0 0 0

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
690 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 47, NO. 5, OCTOBER 2017

x = [ (xc )T ν T ]T . The augmented-state system is then given by white Gaussian with covariances given by
E(wt wsT ) = Qw δts , E(wt vsT ) = Qc δts ,
E(vt vsT ) = Qv δts
Ac (ν) B c (ν)
ẋ = x+ u (7a) (12)
0 0 where Qw , Qc , and Qv are the covariance matrices and δts is
the Kronecker delta function defined by
y = Cc 0 x (7b) 
1, if t = s
where u = uc . Notice that although the system in (3a) and (3b) δts = (13)
is linear, the system in (7a) and (7b) is nonlinear, since the 0, if t = s.
matrices Ac and B c depend on the augmented state x. If we We assume that wt and vs are independent Gaussian random
discretize the system in (7a) and (7b), we obtain the following variables and hence the cross term Qc in (12) is zero. The state
discrete augmented system with additive noise terms estimates can then be computed using the EKF algorithm [34].
xk +1 = AD (ν)xk + BD (ν)uk + wk (8a) An alternative to EKF is to use an UKF. A UKF implements
the unscented transform (UT) [32], and avoids calculating the
yk = CD xk + vk (8b) Jacobian matrices at each time step. Hence, it captures the true
mean and the covariance of the state Gaussian random variable
where wk and vk are the process noise and the measure noise,
to at least second-order accuracy for any nonlinearity. Let us
respectively. As usual, these noise terms are included to model
consider the system in (11a) and (11b). The UKF redefines the
neglected/unmodeled uncertainties.
state vector as xak = [xTk , wkT , vkT ]T and estimates xak recursively.
In the following sections, we estimate the state vector of the
The UT sigma point selection scheme is applied to calculate the
system in (3a) or (7) based on the available data, subject to the
sigma matrix Xka for the augmented state xak .
following constraints:
Although the UKF-based algorithms (joint/dual UKF) are
pi  gi (ν)  pi , i = 1, 2, . . . , 12 (9) expected to have better accuracy, the choice between the joint
estimation and the dual estimation is still not clear, since they
where gi (ν) is the ith element of the vector-valued function g(ν) show different performances when they are applying to different
given by problems. More discussions can be found, for instance, in [32],
[33].
g(ν) = [ 1/ν1 1/ν2 1/ν3 ν4 /ν3 ν5 /(ν2 ν6 )
(ν6 ν7 )/ν1 ν7 ν8 ν9 1/ν10 ν11 1/ν12 ]T . (10) B. Nonlinear State Constraints
Recall that the parameter vector ν to be estimated is con-
Note that some of the parameters in the feedback model, in
strained by the nonlinear inequalities in (9). The Kalman filter-
particular in the neuromuscular system Gnm (s), can be con-
ing constrained state estimation problem has been solved using
sidered to be constants that do not change significantly from
a number of algorithms [40]–[42]. The available approaches for
driver to driver [9], [39]. These parameters will be discussed in
solving linear equality constraint problems include model reduc-
Section VI-A.
tion [43], perfect measurement [44], estimate projection [40],
system projection [45], and soft constraints [46]. The available
III. DRIVER PARAMETER ESTIMATION methods for solving nonlinear equality constraints problems
In this section, we use a joint EKF/UKF and a dual EKF/UKF include Taylor expansion approximation [47], smoothly con-
to estimate the system states and obtain the unknown driver pa- strained Kalman filter [48], moving horizon estimation [49],
rameters. The joint EKF/UKF includes the unknown parameters unscented Kalman filtering [50], and particle filters [51]. In this
into the original state vector and then estimates the states and study, we use the estimate projection algorithm and the first-
the parameters simultaneously. The dual EKF/UKF separates order Taylor expansion approximation method to solve the state
the states and the parameters, so as to estimate the states and the estimation problem with nonlinear inequality constraints in (9).
parameters separately. Geometrically, the idea is to project the unconstrained esti-
mate x̂(k) onto the constraint surface. Mathematically, we solve
A. Nonlinear Kalman Filter the following minimization problem:

The EKF is a classical approach to solve nonlinear estimation x̃k = argmin (x − x̂k )T W (x − x̂k ) (14a)
x
problems. This is achieved by means of linearizing the nonlin-
ear state transition and nonlinear observation models. Let the such that g(x)  b (14b)
discrete system where x̂k and x̃k are the unconstrained estimate and the con-
xk +1 = f (xk , uk , wk ) (11a) strained estimate of the state at the time step k, respectively,
W is the weighting matrix, and g : Rn → Rm is a nonlinear
yk = h(xk , uk , vk ) (11b) vector-valued function. Performing a Taylor series expansion of
(14b) around x̂(k), yields
where wk and vk are the process noise and the measure noise,
respectively, both of which are assumed to be with zero-mean g(x) ≈ g(x̂k ) + g  (x̂k )(x − x̂k ) + · · · , (15)

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
YOU et al.: NONLINEAR DRIVER PARAMETER ESTIMATION AND DRIVER STEERING BEHAVIOR ANALYSIS 691

Fig. 3. Proving ground by the Google map. Fig. 4. Experiment vehicles and some apparatus. 1st row: Fiesta (left), MKS
(medium), F150 (right); second row: power source (left), power converter
(medium), CAN case (right).
and after ignoring higher order terms, we obtain a linear approx-
imation of the constraint inequalities in (14b)
g  (x̂k )x  b − g(x̂k ) + g  (x̂k )x̂k . (16)
The minimization problem (14a) subject to the linear inequal-
ity constraints in (16) can be solved using standard quadratic
programming [11], [52].

IV. FIELD TESTS

Several field tests were conducted to validate the previous
Fig. 5. Illustration of the CAN network on MKS.
driver model. The field tests took place at the ford dearborn
proving ground (DPG) in Michigan during November 2015. The
Ford DPG is about 1750 ms from West end to East end and about
900 m from South end to North end. The width of the double- TABLE I
SHC (CONSTANT VELOCITY); CW=CLOCKWISE, CCW=COUNTER
lane road is about 6 m. Three kinds of tests were conducted. CLOCKWISE
A steering handling course (SHC) test, a fixed-radius circling
(FRC) test, and the public road test (PRT). The SHC and FRC
tests were conducted at zone 1 and zone 2 of the proving ground,
respectively (see Fig. 3). Three vehicles differing in size and
engine power were prepared and were driven by a professional
driver mimicking three different types of drivers having distinct
driving skills (novice, experienced, and racing). This was done
mainly for safety reasons, as untrained novice drivers were not
allowed to use the DPG. Consequently, a natural next step along
this research direction would be the collection of more data
(primarily from untrained novice drivers on the road) in order module (RCM). These ECUs share the data on the HS-CAN bus.
to further corroborate the conclusions of this paper. The INFO-CAN connects to the in-vehicle communications and
In both the SHC test and the FRC test, the driver was re- entertainment system—called the SYNC system, which incor-
quired to maintain the vehicle at a constant velocity throughout porates the global position system (GPS) and the navigation
the road, while in the PRT test, the driver drove freely on a module.
section of a prechosen public road, considering the specific traf- The data collected during the tests were the steering wheel
fic conditions. The proving ground, the experimental vehicles angle, the steering column torque, the yaw rate, and the longi-
and the drivers were provided by the Ford Motor Company. tude and the latitude of the vehicle. The signals of the steering
Fig. 4 shows the experimental vehicles and the main equipment wheel angle and the steering column torque were provided by
used for the tests. All data were collected through a controller the EPAS, the yaw rate signal was provided by the RCM and
area network (CAN) analyzer that interconnected the computer position information was provided by the on-board GPS system.
and the in-vehicle CAN buses. There are two CAN channels, Additional variables such as the vehicle yaw angle, the veloc-
namely, the HS-CAN and INFO-CAN channel, both of which ity/acceleration of the vehicle, the side slip angle, and the road
have a data transfer rate of 500 [kB/s]. The HS-CAN con- curvature were estimated based on the yaw rate and the GPS
nects to most of the regular on-board electronic control units position data. The CAN bus network of the MKS is shown in
(ECU), such as the antiskip braking system (ABS), the electric Fig. 5. The setup of the test conditions for the SHC, FRC, and
power assisted steering (EPAS) system and the restraints control PRT tests are summarized in Table I.

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
692 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 47, NO. 5, OCTOBER 2017

We first converted the GPS coordinates to ECEF coordinates

using the following equations:
a cos φ cos λ a cos φ sin λ a(1 − e2 ) sin φ
x= , y= , z=
χ χ χ
 (17)
where χ = 1 − e2 sin2 φ, a ≈ 6.39 × 106 [m] and e2 ≈
6.69 × 10−3 are the semimajor axis and the first numerical ec-
centricity of the earth, respectively. By performing a Taylor
expansion of (17) about φ and λ and omitting terms higher than
second order, we obtain
We finally rotate the ECEF coordinates to obtain the ENU
coordinates using the following equations:
    ⎛ dx ⎞
de − sin λ cos λ 0
Fig. 6. Illustration on the different coordinate systems. = ⎝ dy ⎠ .
dn − sin φ cos λ − sin φ sin λ cos φ dz
(19)
V. DATA ANALYSIS AND RESULTS The trajectory of the vehicle can be obtained by integrating
In this section, we summarize the data processing step from the ENU coordinates de and dn in (19). The side slip angle β is
the driving tests and we apply the joint EKF/UKF and the dual estimated using the equation [54]
EKF/UKF to estimate the parameters of the assumed driver V 
y
model. Based on the data analysis, a refined driver model is β = arctan −ψ (20)
Vx
proposed to better reproduce the actual steering wheel torque
command of the driver. where Vx and Vy are the longitudinal velocity and the lateral
velocity of the mass center of the vehicle chassis, respectively,
and ψ is the yaw angle. The road curvature ρ is calculated by
A. GPS Data Processing 
Ycog
ρ= 2 )3/2
(21)
Since the road curvature and the side slip angle of the vehicle (1 + Ycog
were not directly measured, we first obtain the missing values by   
where Ycog = ∂Ycog /∂Xcog , Ycog = ∂Ycog /∂Xcog , and Xcog and
processing the GPS data, which are given in the form of latitude
Ycog are the coordinates of the vehicle in the local ENU system.
and longitude. We refer to the method proposed in [53], by
which the GPS coordinates are transformed to local navigation
B. Driver Parameter Identification
coordinates East, North, and Up (ENU) (in this paper, the height
is zero since the vehicle is traveling on the ground). Three useful This section shows the results from the previous driver pa-
coordinate systems used in this transformation are shown in rameter identification and validation procedure, and provides a
Fig. 6, namely, the World Geodetic System 1984 (WGS84), the comparative analysis. Before processing the field test data, we
Earth Centered Earth Fixed (ECEF) system and the ENU system. first implemented the identification approach on a set of data
The WGS84 system expresses the position vector in terms of obtained from CarSim/MATLAB simulation. This was done in
the longitude, the latitude and the height (φ, λ, h) of the vehicle, order to confirm the correctness and limitations of the identifi-
while the ECEF system is in terms of the vehicle Cartesian cation algorithm.
coordinates (x, y, z). The ENU system is represented locally, 1) CarSim Data Processing: The vehicle model used in
which usually works as the navigation coordinate system. the simulation was configured with CarSim 8.0 [55] and was

a cos λ sin φ(1 − e2 ) a sin λ cos φ 1

dx = − 3
dφ − dλ + a cos φ cos λ(−2
χ χ 4
a sin λ sin φ(1 − e2 ) a cos λ cos φ 2
− 7e2 + 9e2 cos2 φ)dφ2 − dφdλ − dλ
χ3 2χ
a sin λ sin φ(1 − e2 ) a cos λ cos φ 1
dy = − dφ + dλ + a cos φ sin λ(−2
χ3 χ 4
a cos λ sin φ(1 − e2 ) a sin λ cos φ 2
− 7e2 + 9e2 cos2 φ)dφ2 − 3
dφdλ − dλ
χ 2χ
a cos φ(1 − e2 ) 1
dz = dφ + a sin φ(−2 − e2 + 9e2 cos2 φ)dφ2 . (18)
χ3 4

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
YOU et al.: NONLINEAR DRIVER PARAMETER ESTIMATION AND DRIVER STEERING BEHAVIOR ANALYSIS 693

TABLE II
CONSTANT PARAMETERS OF THE SYSTEM

m Mass of vehicle 1653 kg

f Distance from center of gravity to front axis 1.402 m

r Distance from center of gravity to rear axis 1.646 m
Ls Distance from center of gravity to far-field visual point 15 m
Iz Moment of inertia of the vehicle 2765 kgm2
Js Moment of inertia of the steering column 0.11 kgm2

TABLE III
DRIVER MODEL PARAMETERS; JEKF=JOINT EKF, DEKF=DUAL EKF,
UB=UPPER BOUND, LB=LOWER BOUND

Parameter JEKF s JEKF JUKF DEKF DUKF UB LB

K a [Nm] 56.56 22.10 21.62 21.29 21.29 100 5

K c [Nm] 19.82 149.87 152.35 151.88 150.96 200 5 Fig. 8. Data, the training curve, and the simulated curve from the Joint/Dual
T L [s] 0.90 0.33 0.34 0.33 0.33 5 0 E-/UKF.
T I [s] 0.48 0.26 0.26 0.26 0.26 0.5 0
T N [s] 0.30 0.18 0.19 0.19 0.20 0.3 0.01
tp [s] 0.19 0.11 0.11 0.11 0.11 0.5 0.01
s [ m ] 3.47 12.06 12.16 12.25 12.07 15 3 from the data, the green dot-dash curve shows the estimation of
K D [Nm] 1.50 0.37 0.27 0.11 0.31 1.5 0.1 the steering wheel torque during the training process, and the
KG −0.41 −0.74 −0.64 −0.79 −0.43 −0.4 −1.5 red solid curve shows the validation result, which is obtained by
T k1 [sec] 1.05 1.50 1.54 1.97 1.57 6 1
T k2 [sec] 5.13 3.82 3.71 3.42 3.81 6 1 using the identified driver model parameters from the training
T k3 [sec] 0.01 0.01 0.01 0.01 0.01 0.03 0.01 process in the simulation. The simulated result agrees well with
the data. The identified driver model parameters are given in the
second column of Table III.
2) Field Test Data Processing: We processed the field test
data using the joint EKF, the joint UKF, the dual EKF, and
the dual UKF separately, so that we can compare the identified
driver parameters obtained from these four different methods.
For instance, we took the set of data from the SHC tests corre-
sponding to the conditions of a “Racing Driver” with a constant
velocity of 45 [mph] in Table I. In each implementation, we
used the first 60% of the data for parameter training and then
Fig. 7. Data, the training curve, and the simulated curve for the steering wheel used the remaining 40% of the data for validation.
torque. By designing the appropriate Kalman filter parameters, such
as the process noise covariance, the measurement noise covari-
initialized with a constant speed of 15 [m/s](54 [km/h]). Other ance, and the initial state covariance matrix, we obtained rea-
vehicle constants can be found in Table II. In addition, we as- sonably good estimation of the parameters. The process noise
sumed a high-adhesion asphalt pavement with a constant friction covariance is considered to be the most critical, and therefore
factor of 0.89 for all simulations. The length and the width of had to be carefully tuned [56], [57]. Fig. 8 illustrates the steer-
the road were configured as 1000 [m] and 6 [m], respectively. ing wheel torque from data, the training curve for each filter,
A path composed of a sequence of straight segments, circular and the simulated output corresponding to the identified model
segments, and clothoids was given as an input. The configured parameters. The green dashed plots in Fig. 8 show how the pre-
road curvature was obtained through a sensor provided by Car- diction of the steering wheel torque at the current time step,
Sim. Since the road curvature data from CarSim are noisy, we provided by the joint/dual E-/UKF based on past data, agrees
applied a first-order lowpass filter with a cutoff frequency at with the current data. After about 1 min the prediction results
2.5 rad/s to eliminate the noise before inputting this signal to get stabilized and agree well with the data.
the driver model. The trajectories of the estimated states (we only show the
After collecting the necessary simulation data, namely, the driver parameters, and each parameter is scaled such that the
steering wheel angle δs , the road curvature ρ, the side slip angle initial value is one) corresponding to the joint EKF are given
β, and the yaw rate r in the input vector uc , and the steering in Fig. 9. The red dash-dot plots in Fig. 8, which are drawn
wheel torque Tdr in the output y c . We then implemented the joint to validate the identified driver parameters, agree well with the
EKF to estimate the driver model parameters. The results are data. Although one sees some difference between the validation
given in Fig. 7. We only show the results of the joint EKF here, results and the data, the results are reasonable, since the param-
since the results given by the other filters were quite similar. In eters of the real driver may change slowly with time. This effect
Fig. 7, the blue dashed curve shows the steering wheel torque is investigated next.

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
694 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 47, NO. 5, OCTOBER 2017

Fig. 9. Time histories of the driver parameters during the training process.
Steady state is reached after 45 s.
Fig. 11. Trajectory of the driver parameters with ±2σ error during the training
process.

Fig. 10. Data, the training curve, and the simulated curve from the Joint UKF.

Fig. 12. Detail of estimate of s along with the 2σ confidence bounds.

C. Driver Model Refinement
Based on the results from the previous section, we refined and are thus, omitted. Fig. 10 indicates that the parameterized
the model by assuming that the process noise for the parameter driver model with time-varying parameters characterizes the
vector ν is colored. To this end, we let driver’s steering behavior much more accurately. This implies
that the parameterized two-point visual driver model architec-
ν̇ = ζ, ζ̇ = ξ (22) ture shown in Fig. 1 is valid, but the parameters are not nec-
essarily constant, and may vary slowly with time. The driver
where ξ is a zero-mean white process noise and ζ is a colored
parameters convergence with time, along with their 2σ-bounds,
process noise with unknown time-varying mean. By discretizing
are shown in Fig. 11. Fig. 12 shows in greater detail the estimate
(22) with a sampling interval Δt, one obtains
(solid red line) and the confidence levels (blue dotted line) for
νk = νk −1 + Δt ζk −1 , ζk = ζk −1 + Δt ξk −1 . (23) s . The plots for the other parameters are similar. From Fig. 11
one observes that, although most of the parameters converge to
If ξk −1 is uncorrelated with ζk −1 , ζk is colored process noise some constants, some of them exhibit drift, specifically, Ka , Kc ,
in the sense that ζk is correlated with itself at different time steps and TL . This behavior accounts for the difference between the
[58] simulation result (red line) and the test data (blue line) shown
E(ζk ζkT−1 ) = E(ζk −1 ζkT−1 + Δt ξk −1 ζkT−1 ) = E(ζk −1 ζkT−1 )  Qζ in Fig. 8. In terms of driver parameter identification, this result
(24) suggests that Ka , Kc , and TL are the most important parameters
where Qζ = 0 is the covariance matrix. For the noise ζ at to track in an online identification scheme.
any two different time steps t and s (t > s) one obtains that
E(ζt ζsT ) = E(ζt−1 ζsT ) = · · · = E(ζs ζsT ) = Qζ , and hence one VI. DRIVER COMPARISON AND ANALYSIS
can summarize the covariances for ζ and ξ as follows: Section V estimates the identified driver parameters using
E(ζt ζsT ) = Qζ , E(ζt ξsT ) = 0, E(ξt ξsT ) = Qξ δts (25) four different nonlinear filters. Our main motivation for param-
eter estimation is to be able to distinguish the different driver
where Qξ is the covariance matrix and δts is given by (13). styles based on the identified driver parameters from experimen-
Equations (22) allow the parameter vector ν to drift with time. tal data. Empirical evidence suggests that one potential strong
We implemented all filters using this model and recorded the distinguishing feature of driver style is the smoothness of the
estimates of ν at each time step. We then performed simulations applied steering command [59], [60]. In order to test this the-
with the time-varying parameters. Fig. 10 shows the results for ory, we first analyzed the driver’s steering behavior according
the Joint UKF case. The results with the other filters are similar, to the identified driver parameters, and we then analyzed the

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
YOU et al.: NONLINEAR DRIVER PARAMETER ESTIMATION AND DRIVER STEERING BEHAVIOR ANALYSIS 695

TABLE IV
DRIVER MODEL PARAMETERS

Parameter Ka Kc TL TI TN tp s KD Kg T k1 T k2 T k3

30 mph Racing 21.7 153.5 0.33 0.26 0.19 0.11 12.1 0.28 −0.66 1.57 3.72 0.013
Experienced 21.9 158.6 0.35 0.28 0.20 0.11 12.1 0.66 −0.40 5.95 3.73 0.013
Novice 17.1 113.5 0.25 0.20 0.15 0.10 8.7 0.37 −0.40 2.92 3.29 0.013
45 mph Racing 21.9 155.8 0.33 0.26 0.19 0.11 12.1 0.30 −0.66 1.53 3.72 0.013
Experienced 21.8 156.8 0.35 0.28 0.19 0.11 12.2 0.35 −0.78 2.31 3.73 0.013
Novice 17.4 121.4 0.27 0.21 0.16 0.08 8.8 0.33 −0.81 5.65 3.21 0.013

driver’s steering behavior by comparing the wavelet transform

of the control signals from different drivers, since it is well-
known that wavelet transform contains information about the
local smoothness of a signal [61].

A. Driver Parameter Analysis

Table IV shows the parameters for the racing, experienced
and novice driver in the SHC tests (MKS vehicle). Since some
of the parameters are varying with time (namely, Ka , Kc , and
TL ), we only show their time-average values in Table IV.
From the tests shown in Table IV, one observes that Kc is
much larger than Ka . This indicates that in the lane-keeping
task the driver pays more attention to θnear than θfar , as expected.
This result may change, however, for a different driving task Fig. 13. Bode plots of G c for the novice, the experienced and the racing
drivers, 45 mph.
[22]. In this section, we wish to compare the experienced driver
steering command with the novice driver steering command for
the same task. The question we wish to answer is whether we quencies of θnear . The near field of view visual distance s for
are able to distinguish between these two (supposedly) distinct the novice driver in Table IV is smaller than that for the expe-
driver styles by analyzing only the driver steering command. rienced and the racing drivers. This does not necessarily imply,
From Table IV, one sees that the parameters Ka , Kc , TL , and TI however, that the larger the preview distance s the better. For
for the novice driver are smaller than that for the experienced instance, Damveld and Happee [62] observed that the driver’s
driver. The anticipatory gain Ka and the compensatory gain Kc compensatory behavior is reduced with increasing preview dis-
represent the attention the driver pays to θfar and θnear , respec- tance (5–100 [m]) and pointed out that, with a preview distance
tively. An increase of Ka leads to dθfar /dt < 0 (oversteering), above a certain point, the drivers no longer minimize the lateral
and the vehicle gets closer to the inner curb of the road. An error but use the additional preview to obtain a smooth path. The
increase of Kc leads to more compensation (dθnear /dt < 0), and preview distance may also depend on the road geometry [63].
the vehicle gets closer to the road centerline. Saleh et al. [30] We analyzed the high-frequency feedback part Gfb =
mentioned that Kc may depend on the driver’s cautiousness Gk1 (1 + Gk2 ) of the model (see Fig. 1), where Gk1 and Gk2
(e.g., the driver avoids driving too close to the border line) and are given by (1). The Bode plots for the three drivers using the
small Kc leads to a great tendency to cut around the bends. The identified parameters are shown in Fig. 14. Fig. 14 shows that
parameters TL and TI define a lead compensation in the com- Gfb has phase lead and high pass properties, and the three Bode
pensatory control path of the driver. A larger TL corresponds plots look quite close to each other. Considering that the in-
to higher compensation rate of θnear (the speed of θnear to reach put signal to Gfb (steering wheel angle) typically does not have
the desired value), but the system will be oscillating if TL is too many high-frequency components, and the low frequencies are
large [30]. TI determines the bandwidth of the frequencies of filtered, only a narrow band of frequencies can effectively pass
θnear to be compensated. Small values of TI mean that the driver through Gfb as feedback to the driver. This result indicates that
compensates all frequencies including the high-frequency noise, the effect of the high-frequency feedback may be small, and
hence leading to an oscillatory system. If TI is large (TI < TL ), hence the visual information is more important to the driver
the bandwidth of the compensatory loop is narrow such that than the steering feel in his/her hands during a lane-keeping
most frequencies of θnear are filtered. task. Fig. 14 also indicates that the parameters in Gfb do not
The Bode plots of the lead compensator Gc are shown in distinguish between the drivers since the Bode plots are close
Fig. 13. One sees that the magnitude of Gc for the novice driver to each other. We may thus be able to fix the parameters in
is the smallest and the center frequency is the highest. This Gfb to represent the steering behaviors of different drivers, as
indicates that the compensatory control of the novice driver is in [9], [22], where KD and Tk1 are considered to be constants
slow and the driver is more likely to compensate the high fre- (i.e., KD = 1, Tk1 = 2.5). It is worth mentioning that, besides

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
696 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 47, NO. 5, OCTOBER 2017

Fig. 16. Steering wheel torque of the racing, experienced and novice driver
(MKS, 45 mph).

Fig. 14. Bode plots of G fb for the novice, the experienced and the racing
drivers (45 mph).

Fig. 15. Plots of K a versus K c for the three types of drivers. Fig. 17. Wavelet transform of T dr of the racing (above), experienced (medium)
and novice (below) driver.

KD and Tk1 in [9], [22], the time delay tp and the parameter TN
steering wheel torques shown in Fig. 16 using the real-valued
in the neuromuscular system are also treated as constants (i.e.,
Daubechies wavelet 3 function with respect to (26). The graphs
tp = 0.151, TN = 0.11).
of the absolute coefficients of the CWT of the steering wheel
Fig. 15 shows the plots of Ka versus Kc . As shown from
torques in Fig. 16 are shown in Fig. 17. The color regions of the
the analysis of the experimental data in Section V-C, the gains
graph indicate the local modulus maxima. The results in Fig. 17
Ka and Kc drift with time. Furthermore, the parameters of the
show that, in the same SHC, the CWT of the control signal of
novice driver change faster and take values in a larger range
the novice driver has more local maxima than the experienced
than both the experienced driver and the racing driver. This
and the racing driver. This may be used to evaluate the perfor-
result may indicate that, at least for a lane-keeping task, the
mance of the steering behavior of the driver. The local maxima
steering behavior of the experienced driver and the racing driver
can be used to detect the position of the local singularities, as
is smoother than the novice driver. To confirm this conjecture,
well as to determine the associated Lipschitz exponents using
in Section VI-B we perform a wavelet analysis of the control
the following theorem [61].
signals of the experienced/racing driver and the novice driver
Theorem 6.1: Suppose that the wavelet ψ(t) is the nth
and compare the two.
derivative of a smooth function, is n times continuously dif-
ferentiable, and has compact support. Let f (t) be a tempered
B. Wavelet Analysis of Driver Steering Torque Command
distribution whose wavelet transform is well defined over [a, b],
In this section, we compare the steering commands of the and let τ0 ∈ [a, b]. Assume that there exists s0 > 0 and a con-
novice, experienced and racing drivers in terms of their fre- stant C, such that for all τ ∈ [a, b] and s < s0 , the modulus
quency characteristics, and local smoothness properties. Recall maxima of W f (s, τ ) belong to the cone defined by
that the continuous wavelet transform (CWT) for a given signal
f at scale s  0 and translation τ ∈ R is written as [64], [65] |τ − τ0 |  Cs. (27)
 +∞ t − τ  Then f (t) is uniformly Lipschitz n in a neighborhood of τ , for
1
W f (s, τ ) = √ f (t)ψ ∗ dt (26) all τ ∈ [a, b], τ = τ0 . Furthermore, f (t) is Lipschitz α (α < n)
s −∞ s
at τ0 , if and only if there exists a constant A such that at each
where ψ ∗ is the complex conjugate of the mother wavelet ψ. modulus maximum (s, τ ) in the cone (27)
Many available wavelet bases can be used, such as Morlet, Paul,
Haar, Daubechies, Coiflets, and Symlets [66], [67]. |W f (s, τ )|  Asα . (28)
Fig. 16 shows the steering wheel torques for the novice, expe- By taking the logarithm of both sides of (28), one obtains
rienced and racing driver, respectively. In order to compare the
frequency content of the signals, we performed a CWT of the log |W f (s, τ )|  log A + α log s. (29)

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
YOU et al.: NONLINEAR DRIVER PARAMETER ESTIMATION AND DRIVER STEERING BEHAVIOR ANALYSIS 697

Fig. 18. Absolute CWT coefficients |W f (s, τ )|.

Fig. 20. Histogram of the Lipschitz exponents α for the novice and experi-
enced driver.

Lipschitz exponents corresponding to the novice driver and the

experienced driver. By comparing the results in Fig. 20, one
observes that the control signal (Tdr ) of the novice driver has
a larger number of singularities, and the Lipschitz exponent α
shows a larger range from −0.1074 to 2.204, with a larger stan-
dard deviation of 0.6876. This result implies that the steering
wheel torque command of the novice driver is more noisy than
Fig. 19. Histogram of the Lipschitz exponents α for the experienced and
racing driver. the experienced driver (see Fig. 16). Furthermore, the distribu-
tion of α exhibits multiple modes, and the first mode on the left
is smaller than that of the experienced driver. These features
The Lipschitz exponent α is therefore determined by the max- observed from the distribution of the Lipschitz exponents may
imum slope of log |W f (s, τ )| on a logarithmic scale. Here, we be used to distinguish the control signals of different drivers and
only perform CWT of the steering wheel torque of the experi- hence classify drivers into different groups.
enced driver and show the process to determine the Lipschitz
exponent. We adopt the Daubechies wavelet 3 that is orthog-
onal, compactly supported and has three vanishing moments, VII. CONCLUSION
by which we can determine Lipschitz exponents α < 3. Fig. 18 This paper adopts the parameterized two-point visual driver
plots the absolute CWT coefficients in the time-scale domain. model to characterize the steering behavior of a driver, and
We find the lines of maxima from Fig. 18 and determine the conducts a series of field tests to investigate the validity of this
positions of the singularities. The singularities are all the points model to predict human driver behavior and driving style. We
on the time axis that the lines of maxima converge to. There have implemented four nonlinear filters, namely, the joint EKF,
may be multiple lines of maxima converging to the points that the joint UKF, the dual EKF, and the dual UKF to estimate the
are close to each other, due to the number of the vanishing parameters based on field test data conducted at Ford’s Dearborn
moments of the mother wavelet or the Lipschitz exponent of the Development Center test facility. The validation results agree
singularity [61]. well with the data. The UKF is considered to be more accurate
Fig. 19 illustrates the histogram showing the distribution of than the EKF in propagating the Gaussian random variables,
the Lipschitz exponents corresponding to the experienced and but the difference is not obvious in this paper. The results of
racing driver. By comparing the results in Fig. 19, one observes our investigation indicate that some of the driver parameters are
that the control signal (Tdr ) of the racing driver has a smaller not exactly constant, but rather vary slowly during a driving
number of singularities. This result indicates that the steering task more that few minutes long. This observation suggests that
wheel torque command of the racing driver is smoother than similar parameterized driver models need to incorporate this
the experienced driver. The distribution of the Lipschitz expo- effect to faithfully represent reality.
nent α of the racing driver shows a smaller minimal, maximal, The main difficulty with the use of the two-point visual driver
and mean value than the experienced driver. This statistical re- model in practice is the difficulty of reliably measuring the two
sult indicates that the singularities of the steering wheel torque visual angles θnear and θfar . In this paper, the road curvature ρref
command of the racing driver are likely to be more irregular is calculated using the GPS data along with the linear estimator
and impulsive than the experienced driver. The racing driver in (2) to obtain the values of θnear and θfar . In general, one will
perhaps tends to sacrifice smoothness locally (smaller Lipschitz need to estimate ρref if it is not readily available [37].
exponents) by making good use of the double-lane road, such The parameters of the driver model provide some interest-
that (s)he could obtain overall better smoothness (fewer sin- ing features to better understand the steering behavior of dif-
gularities) than the experienced driver. The experienced driver, ferent types of drivers. An experienced driver is likely to pay
who was only allowed to drive within a single lane of the road, more attention to both the anticipatory control (θfar ) and the
behaved less aggressively since the mean and minimal value compensatory control (θnear ) than a novice driver. The model pa-
of the Lipschitz exponents are larger than the racing driver. rameters may also depend on the driving task [22]. The wavelet
Fig. 20 illustrates the histogram showing the distribution of the transform provides insights into the driver’s control signal, in

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
698 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 47, NO. 5, OCTOBER 2017

terms of the number and the location of the singularities of the [17] Y. Lin, P. Tang, W. Zhang, and Q. Yu, “Artificial neural network modelling
signal and the distribution characteristics of the associated Lip- of driver handling behaviour in a driver-vehicle-environment system,” Int.
J. Veh. Des., vol. 37, no. 1, pp. 24–45, 2005.
schitz exponents. These can be used to characterize the control [18] R. Hamada et al., “Modeling and prediction of driving behaviors using
signal into different levels of smoothness. Our analysis showed a nonparametric Bayesian method with AR models,” IEEE Trans. Intell.
that the steering wheel torque of an experienced driver has fewer Veh., vol. 1, no. 2, pp. 131–138, Jun. 2016.
[19] C. W. De Silva, Modeling and Control of Engineering Systems. Boca
singularities, and the Lipschitz exponents seem to follow a com- Raton, FL, USA: CRC Press, 2009.
paratively more concentrated distribution. Based on the work of [20] D. L. Wilson and R. A. Scott, “Parameter determination for a Crossover
this paper, a potential next step would be to use machine learning driver model,” Dept Mech. Eng. Appl. Mech., Univ. Michigan, Ann Arbor,
MI, US, Tech. Rep. UM-MEAM-83–17, 1983.
ideas to distinguish the behaviors of the drivers, and to classify [21] L. Saleh, P. Chevrel, F. Claveau, J.-F. Lafay, and F. Mars, “Shared steering
the drivers into distinct categories based on features arising from control between a driver and an automation: Stability in the presence of
the wavelet transform. driver behavior uncertainty,” IEEE Trans. Intell. Transp. Syst., vol. 14,
no. 2, pp. 974–983, 2013.
[22] C. Sentouh, P. Chevrel, F. Mars, and F. Claveau, “A sensorimotor driver
model for steering control,” in Proc. IEEE Int. Conf. Syst. Man Cybern.,
ACKNOWLEDGMENT San Antonio, TX, USA, October 11–14, 2009, pp. 2462–2467.
[23] D. D. Salvucci and R. Gray, “A two-point visual control model of steering,”
The authors would like to thank Mr. D. Starr from Ford Motor Perception, vol. 33, no. 10, pp. 1233–1248, 2004.
[24] J. Steen, H. J. Damveld, R. Happee, M. M. van Paassen, and M.
Company for the excellent work done in the field test. Mulder, “A review of visual driver models for system identification pur-
poses,” in Proc. IEEE Int. Conf. Syst. Man Cybern., Anchorage, AK, USA,
Oct. 9–12, 2011, pp. 2093–2100.
[25] E. Donges, “A two-level model of driver steering behavior,” Human Fac-
REFERENCES tors, vol. 20, no. 6, pp. 691–707, 1978.
[1] National Highway Traffic Safety Administration, “Traffic safety facts [26] M. F. Land and D. N. Lee, “Where do we look when we steer,” Nature,
2014: A compilation of motor vehicle crash data from the fatality anal- vol. 369, no. 6483, pp. 742–744, 1994.
ysis reporting system and the general estimates system,” Dept. Transp., [27] M. F. Land, “The visual control of steering,” Vis. Action, pp. 163–180,
Nat. Highway Traffic Safety Adm., Washington, DC, USA, Tech. Rep. 1998.
DOT-HS-812-261, 2015. [28] H. Neumann and B. Deml, “The two-point visual control model of steering
[2] NHTSA et al., “2015 motor vehicle crashes: Overview,” Traffic Safety – new empirical evidence,” in Digital Human Modeling. Berlin, Germany:
Facts Res. Note, vol. 2016, pp. 1–9, 2016. Springer, 2011, pp. 493–502.
[3] D. Hendricks, J. Fell, and M. Freedman, “The relative frequency of unsafe [29] G. Markkula, O. Benderius, and M. Wahde, “Comparing and validating
driving acts in serious traffic crashes,” Veridian Engineering, Inc., Buffalo, models of driver steering behaviour in collision avoidance and vehicle
NY, USA, Rep. DOT-HS-809–206, 2001. stabilisation,” Veh. Syst. Dyn., vol. 52, no. 12, pp. 1658–1680, 2014.
[4] W. G. Najm, M. D. Stearns, H. Howarth, J. Koopmann, and J. Hitz, [30] L. Saleh et al., “Human-like cybernetic driver model for lane keep-
“Evaluation of an automotive rear-end collision avoidance system,” Dept. ing,” in Proc. 18th World Congr. Int. Fed. Autom. Control, Milano, Italy,
Transp., Nat. Highway Traffic Safety Admin., Washington, DC, USA, Aug. 28–Sep. 2, 2011, pp. 4368–4373.
Tech. Rep. DOT-HS-810-569, 2006. [31] F. Mars, L. Saleh, P. Chevrel, F. Claveau, and J.-F. Lafay, “Model-
[5] G. Li, S. E. Li, and B. Cheng, “Field operational test of advanced driver ing the visual and motor control of steering with an eye to shared-
assistance systems in typical Chinese road conditions: The influence of control automation,” in Proc. Human Factors Ergonom. Soc. Annu.
driver gender, age and aggression,” Int. J. Autom. Technol., vol. 16, no. 5, Meet., vol. 55, no. 1, Las Vegas, NV, USA, Sep. 19–23, 2011,
pp. 739–750, 2015. pp. 1422–1426.
[6] D. H. Weir and D. T. McRuer, “Measurement and interpretation of driver [32] E. A. Wan and R. Van Der Merwe, “The unscented Kalman filter for
steering behavior and performance,” SAE Int., Warrendale, PA, USA, SAE nonlinear estimation,” in Proc. Adapt. Syst. Signal Process. Commun.
Tech. Paper 730098, Feb. 1973. Control Symp., October 1–4, 2000, pp. 153–158.
[7] C. C. MacAdam, “An optimal preview control for linear systems,” J. Dyn. [33] A. Farina, B. Ristic, and D. Benvenuti, “Tracking a ballistic target: Com-
Syst. Meas. Control, vol. 102, no. 3, pp. 188–190, 1980. parison of several nonlinear filters,” IEEE Trans. Aerosp. Electron. Syst.,
[8] C. C. MacAdam, “Application of an optimal preview control for simula- vol. 38, no. 3, pp. 854–867, Jul. 2002.
tion of closed-loop automobile driving,” IEEE Trans. Syst. Man Cybern., [34] G. Chowdhary and R. Jategaonkar, “Aerodynamic parameter estimation
vol. 11, no. 6, pp. 393–399, Jun. 1981. from flight data applying extended and unscented Kalman filter,” Aerosp.
[9] R. Hess and A. Modjtahedzadeh, “A control theoretic model of driver Sci. Technol., vol. 14, no. 2, pp. 106–117, 2010.
steering behavior,” IEEE Control Syst. Mag., vol. 10, no. 5, pp. 3–8, 1990. [35] S. Noth, I. Rañó, and G. Schöner, “Investigating human lane keeping
[10] A. Modjtahedzadeh and R. Hess, “A model of driver steering control through a simulated driver,” in Proc. Driver Simul. Conf., Paris, France,
behavior for use in assessing vehicle handling qualities,” J. Dyn. Syst. Sep. 6–7, 2012, pp. 340–345.
Meas. Control, vol. 115, no. 3, pp. 456–464, 1993. [36] I. Rañó, H. Edelbrunner, and G. Schöner, “Naturalistic lane-keeping based
[11] A. Burgett and R. Miller, “Using parameter optimization to characterize on human driver data,” in Proc. Intell. Veh. Symp., Gold Coast, QLD,
driver’s performance in rear end driving scenarios,” in Proc., Int. Tech. Australia, Jun. 23–26, 2013, pp. 340–345.
Conf. Enhanc. Safety Veh., vol. 2003 May 19–22, 2003, p. 21. [37] C. Sentouh, B. Soualmi, J. C. Popieul, and S. Debernard, “Cooperative
[12] C. Chatzikomis and K. Spentzas, “A path-following driver model with lon- steering assist control system,” in Proc. IEEE Int. Conf. Syst. Man Cybern.,
gitudinal and lateral control of vehicle’s motion,” Forsch. Ingenieurwesen, Manchester, U.K., Oct. 13–16, 2013, pp. 941–946.
vol. 73, no. 4, p. 257, 2009. [38] S. Zafeiropoulos and P. Tsiotras, “Design of a lane-tracking driver steering
[13] D. J. Cole, A. J. Pick, and A. M. C. Odhams, “Predictive and linear assist system and its interaction with a two-point visual driver model,” in
quadratic methods for potential application to modelling driver steering Proc. Amer. Control Conf., Portland, OR, USA, Jun. 4–6, 2014, pp. 3911–
control,” Veh. Syst. Dyn., vol. 44, no. 3, pp. 259–284, 2006. 3917.
[14] S. D. Keen and D. J. Cole, “Application of time-variant predictive con- [39] D. T. McRuer, “Human pilot dynamics in compensatory systems,” , DTIC
trol to modelling driver steering skill,” Veh. Syst. Dyn., vol. 49, no. 4, Document, Wright Air Development Center, Air Research Development
pp. 527–559, 2011. Command, United States Air Force, Wright/Patterson Air Force Base,
[15] M. Flad, C. Trautmann, G. Diehm, and S. Hohmann, “Individual driver Dayton, OH, USA, Tech. Rep. AFFDL-TR-65-15, Jul. 1965.
modeling via optimal selection of steering primitives,” in Proc. World [40] D. Simon and T. L. Chia, “Kalman filtering with state equality constraints,”
Congr., Cape Town, South Africa, August 24–29, 2014, vol. 19, no. 1, IEEE Trans. Aerosp. Electron. Syst., vol. 38, no. 1, pp. 128–136, Jan. 2002.
pp. 6276–6282. [41] N. Gupta and R. Hauser, “Kalman filtering with equality and inequality
[16] I. Kageyama and H. Pacejka, “On a new driver model with fuzzy control,” state constraints,” Oxford University Computing Laboratory, University
Veh. Syst. Dyn., vol. 20, no. sup1, pp. 314–324, 1992. of Oxford, 2007. arXiv:0709.2791.

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.
YOU et al.: NONLINEAR DRIVER PARAMETER ESTIMATION AND DRIVER STEERING BEHAVIOR ANALYSIS 699

[42] D. Simon, “Kalman filtering with state constraints: A survey of lin- [66] C. Torrence and G. P. Compo, “A practical guide to wavelet analysis,”
ear and nonlinear algorithms,” IET Control Theory Appl., vol. 4, no. 8, Bull. Amer. Meteorol. Soc., vol. 79, no. 1, pp. 61–78, 1998.
pp. 1303–1318, 2010. [67] Y.-U. Zhou and J.-Q. Cheng, “Wavelet transformation and its applica-
[43] W. Wen and H. F. Durrant-Whyte, “Model-based multi-sensor data fu- tions,” Acta Phys. Sin., vol. 37, pp. 24–32, 2008.
sion,” in Proc. IEEE Int. Conf. Robot. Autom., Nice, France, May 12–14,
1992, pp. 1720–1726.
[44] L.-S. Wang, Y.-T. Chiang, and F.-R. Chang, “Filtering method for nonlin-
ear systems with constraints,” IEEE Proc.-Control Theory Appl., vol. 149,
no. 6, pp. 525–531, Nov. 2002.
Changxi You received the B.S. and M.S. degrees
[45] S. Ko and R. R. Bitmead, “State estimation for linear systems with state
from the Department of Automotive Engineering,
equality constraints,” Automatica, vol. 43, no. 8, pp. 1363–1368, 2007. Tsinghua University of China, Beijing, China, and the
[46] M. Tahk and J. L. Speyer, “Target tracking problems subject to kinematic
M.S. degree from the Department of Automotive En-
constraints,” IEEE Trans. Autom. Control, vol. 35, no. 3, pp. 324–326,
gineering, RWTH-Aachen University, Aachen, Ger-
Mar. 1990.
many. He is currently working toward the Ph.D. de-
[47] C. Yang and E. Blasch, “Kalman filtering with nonlinear state con- gree at the School of Aerospace Engineering, Georgia
straints,” IEEE Trans. Aerosp. Electron. Syst., vol. 45, no. 1, pp. 70–84,
Institute of Technology, Atlanta, GA, USA.
Jan. 2009.
His current research interests include system iden-
[48] J. De Geeter, H. Van Brussel, J. De Schutter, and M. Decréton, “A smoothly
tification, aggressive driving maneuver, and control
constrained Kalman filter,” IEEE Trans. Pattern Anal. Mach. Intell.,
of (semi)autonomous vehicle.
vol. 19, no. 10, pp. 1171–1177, Oct. 1997.
[49] H. Michalska and D. Q. Mayne, “Moving horizon observers and observer-
based control,” IEEE Trans. Autom. Control, vol. 40, no. 6, pp. 995–1006,
Jun. 1995.
[50] B. O. S. Teixeira et al., “Unscented filtering for interval-constrained
nonlinear systems,” in Proc. IEEE Conf. Decis. Control, Canxun, Mexico,
Dec. 9–11, 2008, pp. 5116–5121.
[51] D. Crisan and A. Doucet, “A survey of convergence results on particle
Jianbo Lu received his B.S. degree in mechani-
filtering methods for practitioners,” IEEE Trans. Signal Process., vol. 50,
cal engineering from the Central South University,
no. 3, pp. 736–746, Mar. 2002.
Changsha, China, and the M.S. degree in mechani-
[52] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.: cal engineering from Arizona State University, and
Cambridge Univ. Press, 2004.
his Ph.D. degree in aeronautics and astronautics from
[53] S. P. Drake, “Converting GPS coordinates [φ, λ, h] to navigation coordi-
Purdue University. He is currently a Technical Expert
nates (ENU), Electr. Surveillance Res. Lab., Edinburgh, South Australia,
in advanced vehicle controls at Ford Motor Company,
Australia, Tech. Rep. DSTO-TN-0432, Apr. 2002. Dearborn, MI, USA. He holds more than 100 U.S.
[54] H. Zhao and H. Chen, “Estimation of vehicle yaw rate and side slip angle
patents and numerous pending patent applications,
using moving horizon strategy,” in Proc. 6th World Congr. Intell. Control
and has published more than 70 journal and confer-
Autom., Dalian, China, Jun. 21–23, 2006, vol. 1, pp. 1828–1832.
ence articles. His research interests include automo-
[55] CarSim Educational Software. Mechanical Simulation Corporation, Ann
tive controls and sensing, adaptive vehicle systems, driver assistance systems,
Arbor, MI, USA, Jan. 2000.
smart mobility, and semiautonomous and autonomous systems.
[56] M. Saha, B. Goswami, and R. Ghosh, “Two novel costs for determining
Dr. Lu received the Henry Ford Technology Reward twice.
the tuning parameters of the Kalman filter,” 2011. arXiv:1110.3895.
[57] M. Ananthasayanam, “Kalman filter design by tuning its statistics or
gains?” in Proc. Int. Conf. Data Assimilation, Mumbai, India, Jul. 13–15,
2011.
[58] D. Simon, Optimal State Estimation: Kalman, H ∞ , and Nonlinear Ap-
proaches. New York, NY, USA: Wiley, 2006.
[59] Y. L. Murphey, R. Milton, and L. Kiliaris, “Driver’s style classification
using jerk analysis,” in Proc. IEEE Workshop Comput. Intell. Veh. Veh.
Panagiotis Tsiotras (SM’02) received the Eng.Dipl.
Syst., 2009, pp. 23–28. degree in mechanical engineering from the National
[60] T. Flash and N. Hogan, “The coordination of arm movements: an ex-
Technical University of Athens, Athens, Greece, the
perimentally confirmed mathematical model,” J. Neurosci., vol. 5, no. 7,
M.S. degrees in aerospace engineering from Virginia
pp. 1688–1703, 1985.
Tech, Blacksburg, VA, USA, and in mathematics
[61] S. Mallat and W. L. Hwang, “Singularity detection and processing with from Purdue University, Lafayette, IN, USA, and the
wavelets,” IEEE Trans. Inf. Theory, vol. 38, no. 2, pp. 617–643, Mar. 1992.
Ph.D. degree in aeronautics and astronautics from
[62] H. Damveld and R. Happee, “Identifying driver behaviour in steering:
Purdue University.
Effects of preview distance,” in Proc. Measuring Behav., Utrecht, The He is currently the Dean’s Professor in the School
Netherlands, Aug. 28–31, 2012, pp. 44–46.
of Aerospace Engineering, and the Associate Director
[63] Y. Hassan and T. Sayed, “Effect of driver and road characteristics on
of the Institute for Robotics and Intelligent Machines
required preview sight distance,” Can. J. Civil Eng., vol. 29, no. 2,
at the Georgia Institute of Technology, Atlanta, GA, USA. His research interests
pp. 276–288, 2002. are in optimal control, nonlinear control, and vehicle autonomy.
[64] A. Grossmann, R. Kronland-Martinet, and J. Morlet, “Reading and under-
Dr. Tsiotras has been in the Editorial Boards of the IEEE TRANSACTIONS
standing continuous wavelet transforms,” in Wavelets. Berlin, Germany:
ON AUTOMATIC CONTROL (2009–2011), the IEEE CONTROL SYSTEMS MAGA-
Springer, pp. 2–20, 1989.
ZINE (2003–2009), and the AIAA Journal of Guidance, Control, and Dynamics
[65] B. Jawerth and W. Sweldens, “An overview of wavelet based multireso- (1999–2007). He received the NSF CAREER Award. He is a Fellow of the
lution analyses,” SIAM Rev., vol. 36, no. 3, pp. 377–412, 1994.
AIAA.

Authorized licensed use limited to: Zhejiang University. Downloaded on September 06,2024 at 02:19:07 UTC from IEEE Xplore. Restrictions apply.