2022 7th International Conference on Mechanical Engineering and Robotics Research

Path Tracking Methods Employed in Lane

Centering Systems
2022 7th International Conference on Mechanical Engineering and Robotics Research (ICMERR) | 978-1-6654-9051-1/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICMERR56497.2022.10097807

Musa Nurullah Yazar, Zülfikar Ali Erbudak, Alperhan Bay, İlkay Dost, Abdurrahim Semiz
AVL Research and Engineering Center, Türkiye
{musa.yazar, ali.erbudak, alperhan.bay, ilkay.dost, abdurrahim.semiz}@avl.com

Abstract—In this paper, a simulation environment with lane Geometric methods can provide efficient performance in a sim-
detection system emulator has been developed in order to com- ple structure employing the geometric models of the vehicle
pare the performance of path tracking methods. Path tracking and road. The Stanley, mostly used geometric method, was
problem is handled in local coordinate frame employing the
lane detection system emulator outputs. By performing this developed by the winning team in the DARPA competition in
environment, a detailed and proper comparison of geometric and 2005 [1]. Pure Pursuit method [2] is also one of the most used
model based path tracking methods, mostly used in the literature, geometric based tracking approaches. Both of these methods
are executed with respect to different speeds and road curvatures. have advantages and disadvantages, it should be taken into
The performance results show that the compared path tracking account when selecting the proper method with respect to
methods do not work well enough for all the determined perfor-
mance criteria and the methods have complementary features. driving scenario. Pure Pursuit performs reasonably well and is
These complementary features show that path tracking methods quite robust to large cross-track errors and discontinuous paths.
can be employed in a hybrid structure to achieve a generally Stanley method outperforms Pure Pursuit method in the case of
better performance results for the speed range (0-130 [km/h]) aggressive turn and high velocities. However, Stanley method
where autonomous driving is enabled. In addition, the obtained is less robust to large cross-track errors and non-smooth
simulation results reveal the necessity of adaptive adjustment of
the gain parameters of the path tracking methods with respect paths. Even though geometric methods generally overcome
to the road curvature and vehicle speed. disturbance inputs, their performance is insufficient in some
Index Terms—autonomous vehicles, lane centering, path track- driving scenarios. Researchers have recently developed hybrid
ing, stanley, purepursuit, lqr methods in which two algorithms are used together to improve
the weaknesses of the individual methods [3]–[5]. Another
I. I NTRODUCTION approach is utilizing machine learning algorithms to choose
In recent years, an intensive research effort has focused the parameters of the methods with respect to path curvature,
on developing advanced driving assistant systems (ADAS) cross-track error and longitudinal velocity [6]–[8].
and autonomous driving, Cruise control (CC) and adaptive In model-based methods, path tracking problem is handled
cruise control (ACC) have been become widely used features in the classical feedback control loop by using the vehicle’s
in today’s vehicles. Beyond such control functions, with the kinematic or dynamic models. A comprehensive analysis and
development of longitudinal and lateral control systems, lane comparison of different feedback structures were performed
keeping (LKA) and lane changing assistant (LCA) systems utilizing the vehicle kinematic and dynamic model [9]. Mod-
have also been made commercially available by many automo- eling errors cause sub-optimal control performance and may
tive manufacturers. LKA and LCA systems require en efficient cause the unstability. The cross-track error and hence tracking
path tracking algorithm and hence lateral and longitudinal performance both suffer from modeling errors, especially at
controller design. There exist many methods in the literature high vehicle speeds. Dynamic model based optimal control
with different complexity and whose performances may vary method may overcome geometric methods in some speed
depending on the driving scenario and conditions. regions. It should be noted that the linearized dynamic model
The purpose of the path tracking methods is to minimize is only valid within a certain speed range. In the model
angular and cross-track error between the path and the vehicle, predictive control, linearized models for different speed ranges
and maintain lateral acceleration below a certain value in order are employed in a certain prediction period and a better
to provide driving comfort. The mathematical model of the tracking performance can be achieved [10].
vehicle to be controlled can be obtained kinematically and In this paper, the geometric and model-based optimal control
dynamically. However, the kinematic model is not appropriate methods employing in the lane centering system are analyzed
for control applications due to the increase in dynamic effects and the detailed comparison of the algorithms are carried
at high speeds. The vehicle dynamic model, on the other hand, out in the simulation environment. The developed algorithms
can be handled separately as lateral and longitudinal dynamics, can be evaluated directly in the real environment thanks
and the controller design can be performed separately. The to the creation of realistic simulation environment. In the
connection between the two dynamics cannot be neglected, second section, the installation of the lane detection system
especially in sudden maneuvering movements at high speeds and the geometric model of the road are presented. Third
and on curvature roads. section introduces mathematical models of the path tracking

978-1-6654-9051-1/22/$31.00 ©2022 IEEE 13

Authorized licensed use limited to: ULAKBIM UASL - YILDIZ TEKNIK UNIVERSITESI. Downloaded on January 31,2025 at 16:58:28 UTC from IEEE Xplore. Restrictions apply.
 methods to be used in comparison. Fourth section presents the lateral dynamic model. The position and orientation of the
simulation results and in the last section, a conclusion is made vehicle are computed for each sampling time and they are
and future studies are suggested. given as an input for Vehicle Lane Detection Camera Coordi-
nate Transformation block which computes the homogeneous
II. S YSTEM M ODEL
transformation matrix. This innovative simulation framework,
Figure 1 shows the position of camera in the lane detection built within the scope of the study, provides more realistic
system on the vehicle and field of view of the camera-based environment to test the performance of the lane tracking
detection system. Vision range of the lane detection system is systems where global map is not available and hence high
determined as 80 [m] under ideal circumstances. However, it accuracy localization is not possible for autonomous driving
can vary depending on environmental factors. software.

Fig. 2. Lane detection system.

Fig. 1. Lane detection system.

III. PATH T RACKING M ETHODS

Figure 2 illustrates the block diagram of the simulation
environment. Global map of the road where the performance A. Stanley Method
of the path tracking algorithms will be compared is given The Stanley method was first developed in 2005 for Stan-
as an input to the system. The waypoints coming from the ford University’s autonomous vehicle, named Stanley, for
global map are transformed to the local coordinate frame the DARPA competition. The Stanley method’s fundamental
by computing the homogeneous transformation matrix with parameters are shown in Figure 3. Equation 2 provides the
respect to position and orientation of the lane detection calculation for the steering angle that minimizes both the
camera on the windshield of the vehicle. Vision range (80 cross-track error and the yaw error. The cross-track error is
meters) representing by closest waypoints to the vehicle in determined by taking the lateral distance between center of
the local coordinate frame is given as input for the lane vehicle’s front axle and the closest point to center lane. The
detection emulator. The lane detection emulator computes the yaw error is calculated by taking the difference between the
coefficients of the third-order polynomial across the points by yaw angle of the vehicle and the heading angle of the road.
utilizing the least-square technique. Lane center polynomial is Equation 2 shows that higher steering angles are calculated
given in Equation 1. In the local coordinate frame x and y at low vehicle speeds whereas smaller steering angles are
represents the longitudinal and lateral directions respectively. calculated as the vehicle speed increases.
Lane polynomial coefficients are obtained instantaneously for
each sampling time. In the lane detection system emulator,
disturbances can be added to the outputs to represent the noisy
data coming from real environment.

ycenter = c0 + c1 x + c2 x2 + c3 x3 (1)
The path tracking method takes the lane polynomial coef-
ficients as an input and computes the required steering angle
which minimizes the tracking errors by using the geometric
models of the vehicle and road. Additionally, cross-track error,
yaw error and steering wheel angular velocity are computed Fig. 3. Stanley method.
as the performance outputs. Steering angle and reference
velocity which is given in a certain range to measure the ker (t)
tracking performance are provided as an input for vehicle δ = ψe (t) + tan−1 ( ) (2)
vx (t)

14
Authorized licensed use limited to: ULAKBIM UASL - YILDIZ TEKNIK UNIVERSITESI. Downloaded on January 31,2025 at 16:58:28 UTC from IEEE Xplore. Restrictions apply.
 B. Pure Pursuit Method
Another widely used geometric-based control method is
the Pure Pursuit method. The parameters used in the Pure
Pursuit method are shown in Figure 4. The target point on the
imaginary lane in the middle of the road lanes (left and right) is
found according to the specified lookahead distance parameter.
Lookahead distance is determined with respect to the vehicle
speed. Using the geometric relationships shown in Figure 4,
the Pure Pursuit method determines the steering angle that
drives the vehicle to approach the target point. Equation 3 to
7 specify the steps for calculating the steering angle based on
the cross-track error with geometric relations.
Fig. 5. Dynamic bicycle model.

In the simplified dynamic model of the vehicle, only the

yaw angle, which is rotation around the z axis, is considered.
The roll and pitch angles are neglected. Moments acting on
the vehicle around the z-axis are given in Equation 12.

ℓf (Fyf cos(δ) + Fxf sin(δ)) − ℓr Fyr = Iz ψ̇ (12)

Iz represents the moment of inertia of the vehicle about the
z axis in Equation 12. The slip angles of the front and rear
wheels are given in Equation 13 and 14.
Fig. 4. Pure Pursuit method.
vy + ℓf ψ̇
αf = tan−1 ( )−δ (13)
vx
ℓd R vy − ℓr ψ̇
= π (3) αr = tan−1 ( ) (14)
sin(2α) sin( 2 − α) vx
ℓd R Lateral forces acting on the vehicle in proportion to the
= (4)
2sin(α)cos(α) cos(α) slip angle of the wheels and providing the vehicle’s lateral
ℓd movement can be expressed as follows by depending on the
R= (5)
2sin(α) friction force between the road and the tire and the cornering
L 2Lsin(α) stiffness parameters (cf and cr ) that vary according to the tire
δ = tan−1 ( ) = (6) characteristics.
R ℓd
er
sin(α) = (7)
ℓd Fyf = −cf αf (15)
L = ℓf + ℓr (8)
Fyr = −cr αr (16)
2Le 2L
δ = tan−1 ( 2 ) ≈ 2 e (9) The dynamic model of the system is obtained by applying
ℓd ℓd
the force equations in Equations 15 and 16 into Equations 11
The Pure Pursuit method acts as a proportional controller as and 12.
shown by Equation 9. As mentioned previously, the parameter
ℓd in the equation can be chosen with respect to the speed. In −cf αf cos(δ) − cr αr
Equation 10, K is the adjustable speed gain coefficient and d v̇y = − vx ψ̇ (17)
m
is the adjustable constant variable. −ℓf cf αf cos(δ) + ℓr cr αr
ψ̇ = (18)
Iz
ℓd = Kvx + d (10)
The system’s linearized model under the assumption of
C. Dynamic Model Based Optimal Control small angles is given in Equation 19.
In Figure 6, the illustration of the forces acting on a vehicle
on a simplified bicycle model is given, Equation 11 specifies −(cf + cr ) (ℓr cr − ℓf cf ) cf
v̇y = vy + [ − vx ]ψ̇ + δ (19)
the translational forces acting on the vehicle in the y direction. mvx mvx m
2 2
ℓr cr − ℓf cf ℓf cf + ℓr cr ℓf cf
ψ̇ = vy − ψ̇ + δ (20)
Fyf cos(δ) + Fxf sin(δ) + Fyr = m(v̇y + vx ψ̇) (11) Iz vx Iz vx Iz

15
Authorized licensed use limited to: ULAKBIM UASL - YILDIZ TEKNIK UNIVERSITESI. Downloaded on January 31,2025 at 16:58:28 UTC from IEEE Xplore. Restrictions apply.
 Since the lane tracking system uses the vehicle local coor- The eigenvalues of the closed loop system (Ac = A − B1 K)
dinate frame, it is essential to represent the system model with must be adjusted by using gain parameter K to obtain desired
the relationships between the road and the vehicle. The local performance outputs. It is very important in terms of solution
coordinate frame in Figure 6 is used to define the dynamic flexibility to use the optimal control method approach instead
bicycle model. The lane centering algorithm is aimed to reduce of manually adjusting the eigenvalues of the system. In the
the error between ψp , which represents the slope angle of optimal control approach, cost function determines the per-
the road at the location of the vehicle mass center closest formance outputs. Since the cost function includes quadratic
to the middle of the lane, and ψ, which represents the vehicle functions, it is called as Linear Quadratic Regulator (LQR) in
orientation.The orientation error is given as the literature. Cost function is given in Equation 26.
1 ∞ T
Z
ψe = ψ − ψp (21) J= x (t)Qx(t) + δ(t)T Rδ(t) dt (26)
2 0
In Equation 26, Q and R are weight matrices related with
states and inputs respectively. If Equation 26 is rearranged
with δ = −Kx;
1 ∞ T
Z
J= x (t)(Q + K T RK)x(t) dt (27)
2 0
If we define P matrix equals to (Q + K T RK), P matrix
becomes a constant matrix consisting of Q and R weight
matrices as well as the K state feedback gain matrix.
d T
(x P x) = −xT (Q + K T RK)x (28)
dt
Fig. 6. Dynamic bicycle model in local coordinate frame. If K matrix is chosen to satisfy Equation 28 and it is
The acceleration and change of the cross-track error are assumed that the states go to zero as time goes to infinity,
given as the cost function is in Equation 29 depends only on the P
matrix and the initial conditions. Thus the cost function can
ër = v̇y + vx ψ̇e (22) reach its minimum value.
ėr = vy + vx ψe (23)
1 ∞ d T
Z
1
The equations given above is used in the linearized model of J =− (x P x)dt = xT (0)P x(0) (29)
2 0 dt 2
the system and rearranged with repsect to the relations between
The main issue here is to determine the gain matrix (K) that
the vehicle and road to obtain state space model of the system
satisfies Equation 28. If we apply the derivative in Equation 28
given in Equation 24.
and replace it in Equation 25 by applying K = R−1 B T P , the
 0  Riccati equation is obtained. The P matrix can be computed

ėr 1 0 0 er

by solving Riccati equation utilizing numerical methods or
−(cf +cr ) cf +cr ℓr cr −ℓf cf
 ër   0 mvx m mvx
  ė 
 r the Schur matrix decomposition method. Finally, the optimal
 =
0 0 0 1
ψ̇e    ψ 
 e state feedback gain is found by using the P matrix in the
−(ℓ2f cf +ℓ2r cr )
ψ̈e 0
ℓr cr −ℓf cf ℓf cf −ℓr cr ψ̇e K = R−1 B T P equation.
Iz vx Iz Iz vx
 

0
 0 IV. S IMULATIONS
 cmf   ℓr cr −ℓf cf − vx 
 mvx The system model shown in Figure 2 is built in Matlab
+
 0 δ +   ψ̇p
 
 0  Simulink environment. Path tracking methods are evaluated
ℓf cf −(ℓ2f cf +ℓ2r cr )
Iz using the same global map and vehicle parameters and their
Iz vx
(24) performance results are compared.
Figure 7 shows the performance outputs of the path tracking
Equation 24 is in the state space model form defined as
methods. Curvature data (κ = R1 ) of the road belongs to the
ẋ = Ax + B1 δ + B2 ψ̇p . In the state space representation, state
global map is given in the bottom of Figure 7. Regarding
vector x = [er ėr ψe ψ̇e ]. The input matrix and residual
the simulation results, the cross-track and yaw error remain
matrix are represented by B1 and B2 respectively. δ and ψ̇ are
acceptable in the regions (between 150 and 250 meters)
provided to the system as inputs. If the controllability matrix is
where the curvature of the road increases whereas the vehicle
full rank, the system can be controlled using full state feedback
speed decreases. However, it is observed that the performance
where δ = −Kx. In this case, the state space model of the
outputs are significantly distorted in the regions where both
closed loop system given as
curvature and vehicle speed increases (between 850 and 950
ẋ = (A − B1 K)x + B2 ψ̇p (25) meters). It can be seen that cross-track error has exceeded

16
Authorized licensed use limited to: ULAKBIM UASL - YILDIZ TEKNIK UNIVERSITESI. Downloaded on January 31,2025 at 16:58:28 UTC from IEEE Xplore. Restrictions apply.
 and yaw errors and steering-wheel velocity which are the
performance outputs of the system. Test scenarios include
various speed combinations and road curvatures. In the
results, it has been determined that the Pure Pursuit method
performs better in low speeds and steering angle changes
are more smooth. Whereas it makes high cross-track error
in curvature regions of the test road especially in high
speeds. The Stanley algorithm performs the best in terms
of minimizing the cross-track error. However, the high peak
values in steering angle changes significantly threaten safety
and driving comfort. Considering the dynamics around the
rolling axis are neglected in the simulations, it is clear to see
that improvements have to be made for the sudden steering
angle changes during implementation in real environment.
Future research will focus on machine learning based
methods that provide adaptive change of the tracking method
parameters with respect to road curvature and vehicle speed.
Additionally, techniques that will allow these methods to
be used in a hybrid structure can be employed to improve
tracking performance.

R EFERENCES
Fig. 7. Simulation results.
[1] S. Thrun, M. Montemerlo, H. Dahlkamp, D. Stavens, A. Aron, J. Diebel,
P. Fong, J. Gale, M. Halpenny, G. Hoffmann et al., “Stanley: The robot
that won the darpa grand challenge,” Journal of field Robotics, vol. 23,
above the acceptable limits at high speeds especially for no. 9, pp. 661–692, 2006.
the Pure Pursuit method. Stanley and LQR method performs [2] O. Amidi and C. E. Thorpe, “Integrated mobile robot control,” in Mobile
more acceptable cross-track error throughout the simulation. Robots V, vol. 1388. SPIE, 1991, pp. 504–523.
[3] M. Cibooglu, U. Karapinar, and M. T. Söylemez, “Hybrid controller
However, instantaneous steering angle changes threaten safety approach for an autonomous ground vehicle path tracking problem,”
and driving comfort. In Figure 8, the steering angle changes in 2017 25th Mediterranean Conference on Control and Automation
are given for each path tracking algorithm. (MED). IEEE, 2017, pp. 583–588.
[4] E. Seo, S. Lee, G. Shin, H. Yeo, Y. Lim, and G. Choi, “Hybrid tracker
based optimal path tracking system of autonomous driving for complex
Steering wheel angular velocity (rad/s)
road environments,” IEEE Access, vol. 9, pp. 71 763–71 777, 2021.
6 [5] Y. Huang, Z. Tian, Q. Jiang, and J. Xu, “Path tracking based on improved
PurePursuit
Stanley
pure pursuit model and pid,” in 2020 IEEE 2nd International Conference
4 Lqr on Civil Aviation Safety and Information Technology (ICCASIT. IEEE,
2020, pp. 359–364.
[6] B. Ağın and M. N. Yazar, “Improving path tracking performance for
2
autonomous driving with reinforcement learning,” Otomatik Kontrol
Ulusal Kongresi, 2022.
rad/s

0 [7] A. AbdElmoniem, Y. T. Afif, S. A. Maged, M. Abdelaziz, and S. Ham-

mad, “Adaptive pure-pursuit controller based on particle swarm opti-
-2
mization (pso-pure-pursuit),” in 2021 16th International Conference on
Computer Engineering and Systems (ICCES). IEEE, 2021, pp. 1–6.
[8] M.-W. Park, S.-W. Lee, and W.-Y. Han, “Development of lateral control
-4 system for autonomous vehicle based on adaptive pure pursuit algo-
rithm,” in 2014 14th International Conference on Control, Automation
-6 and Systems (ICCAS 2014). IEEE, 2014, pp. 1443–1447.
100 200 300 400 500 600 700 800 900 1000
[9] A. D. Luca, G. Oriolo, and C. Samson, “Feedback control of a
meters
nonholonomic car-like robot,” Robot motion planning and control, pp.
171–253, 1998.
Fig. 8. Steering wheel angular velocity. [10] E. Kim, J. Kim, and M. Sunwoo, “Model predictive control strategy
for smooth path tracking of autonomous vehicles with steering actuator
dynamics,” International Journal of Automotive Technology, vol. 15,
V. C ONCLUSION no. 7, pp. 1155–1164, 2014.

This paper presents a comparison of path tracking methods

employed in lane centering systems. The tracking problem
is handled in local coordinate frame, and a lane detection
emulator suitable for the real system is developed and
integrated into the simulation environment. Geometric and
model-based algorithms are compared over the cross-track

17
Authorized licensed use limited to: ULAKBIM UASL - YILDIZ TEKNIK UNIVERSITESI. Downloaded on January 31,2025 at 16:58:28 UTC from IEEE Xplore. Restrictions apply.