Dynamic Speed Adaptation For Path Tracking Based o PDF
Dynamic Speed Adaptation For Path Tracking Based o PDF
Article
Dynamic Speed Adaptation for Path Tracking Based
on Curvature Information and Speed Limits †
Citlalli Gámez Serna *,‡ and Yassine Ruichek ‡
Le2i FRE2005, CNRS, Arts et Métiers, University Bourgogne Franche-Comté, UTBM, F-90010 Belfort, France;
[email protected]
* Correspondence: [email protected]; Tel.: +33-(0)38-458-3441
† This paper is an expanded version of our paper published in Gámez Serna, C.; Lombard, A.; Ruichek, Y.;
Abbas-Turki, A. Gps-based curve estimation for an adaptive pure pursuit algorithm. In Proceedings of the
15th Mexican International Conference on Artificial Intelligence, Cancún, Mexico, 23–29 October 2016.
‡ These authors contributed equally to this work.
Abstract: A critical concern of autonomous vehicles is safety. Different approaches have tried to
enhance driving safety to reduce the number of fatal crashes and severe injuries. As an example,
Intelligent Speed Adaptation (ISA) systems warn the driver when the vehicle exceeds the
recommended speed limit. However, these systems only take into account fixed speed limits
without considering factors like road geometry. In this paper, we consider road curvature with
speed limits to automatically adjust vehicle’s speed with the ideal one through our proposed
Dynamic Speed Adaptation (DSA) method. Furthermore, ‘curve analysis extraction’ and ‘speed limits
database creation’ are also part of our contribution. An algorithm that analyzes GPS information
off-line identifies high curvature segments and estimates the speed for each curve. The speed
limit database contains information about the different speed limit zones for each traveled path.
Our DSA senses speed limits and curves of the road using GPS information and ensures smooth
speed transitions between current and ideal speeds. Through experimental simulations with different
control algorithms on real and simulated datasets, we prove that our method is able to significantly
reduce lateral errors on sharp curves, to respect speed limits and consequently increase safety and
comfort for the passenger.
Keywords: speed adaptation; speed limits; curve speed; path tracking; lateral error; road safety;
autonomous vehicle
1. Introduction
Autonomous vehicle navigation has been a challenging field of study where today the main
goal is to provide safety. In order to accomplish that, vehicle control has to perform accurate path
tracking; in other words, it should minimize the lateral distance between the vehicle’s position and the
defined path.
In real driving scenarios, speeding is one of the main causes for traffic accidents [1], which has
been considered for Advance Driver Assistance Systems (ADAS). The correlation between speeding
and lateral errors is positive, and as a consequence, safety-related issues will depend on vehicle’s
speed control.
Intelligent Speed Adaptation (ISA) systems have been proven in several studies [2–5] to reduce
accidents by respecting speed limits. However, there are several other factors that need to be considered
for precise speed control. One of the major ones involves road geometry.
Analyzing road structure will permit the identification of straight and curved segments
ahead while driving, allowing the lateral control system to adjust the vehicle’s speed accordingly.
Curve detection is of vital importance since the crash rate is at least 1.5-times higher than in
tangent (straight) segments [6]. In addition, when a curve becomes sharper, the number of accidents
increases [7,8]. Hence, different studies have focused on analyzing how processing road structure [9–11]
may benefit control systems.
Using speed limit information, systems like ISA will inform the driver when exceeding speed
limits. On the other hand, approaches considering curvature information [9,10] will emit warnings
when approaching the curve too fast. In any case, control systems require precise knowledge of
the vehicle’s global position. This location information is usually provided by Global Positioning
Systems (GPS), but its main drawback is that it suffers from big positioning errors. These errors are
caused by different factors like blockage, multipath, etc., specially in urban environments. In order
to overcome these big errors, different solutions proposed the use of additional sensors such stereo
vision, IMU, radar or LiDAR [11–14]. For example, vision-based solutions together with digital maps
are capable of estimating the ego motion of the vehicle and correct the vehicle’s global position using a
map-matching algorithm [13]. In [14], the use of cameras, an accurate digital map, GPS and inertial
measurements improve the ego-localization of the vehicle using a variant of the Kalman filter [15].
The main limitation of vision approaches is visibility, since shadows, highlights, occlusions or weather
conditions affect the accuracy of data analysis. Moreover, IMU sensors are able to estimate the current
position of the vehicle. Fusing GPS and IMU with the typically Kalman filter [15] provides absolute
position and orientation, even if GPS data are not available all of the time [16].
Road map databases (GIS maps) together with GPS data provide accurate road geometry and
localization [6,17]. An approach to analyze road geometry identifying critical curves where accidents
occur frequently was proposed in [6]. Wang et al. [17] calculated the road ahead of the vehicle through
a flexible road model. Nevertheless, GIS information is not always available, and GPS is starting to be
accessible or will be soon for the general public. For this reason, the choice of GPS data for tracking
approaches is convenient. Our work is based on Real-Time Kinematics GPS (RTK-GPS), which has the
highest absolute position accuracy (up to a few centimeters), to implement an automatic approach that
goes beyond warnings for adjusting the vehicle speed depending on the upcoming road characteristics,
e.g., curves or speed limit zones.
This paper proposes a Dynamic Speed Adaptation (DSA) method for control systems,
which together with speed limits database creation and curve extraction, adjusts automatically the
vehicle’s speed (see Figure 1). These two approaches, speed limits database creation and curve
extraction, are performed off-line including the preprocessing step. Their output provides the main
parameters to be analyzed by the speed negotiation algorithm of DSA in order to compute the ideal
speed. The identification of sharp curves is performed using GPS positions to define segments that
belong to a curve or a tangent line. This is in order to estimate the convenient driving speed for
the detected high curvature segments. The speed limit database is created for the traveled paths
identifying the positions where speed limits change. As the tracking performance is evaluated through
lateral errors, our DSA is tested in simulations with different steering control algorithms. On the
modeling level, our proposed work would:
Figure 1. Overall workflow of the vehicle path tracking system. The blue dotted bounding box
highlights our contributions.
The rest of this paper is organized as follows. Section 2 presents a review of related works.
Section 3 provides an overview of the steering control algorithms that we used to test our approach.
Section 4 explains our DSA method. In Section 5, we discuss and analyze the results obtained with
and without considering automatic speed adaptation in the different steering control algorithms.
The last section presents conclusions and highlights the main contributions to continue with future
work directions.
2. Literature Overview
As our DSA method estimates the speed based on either curvature information or speed limits,
works related to these two problems are considered.
Lehtonen et al. [26] examined how drivers anticipate their eyes look-ahead fixations on curves to
provide visual guidance for steering. Zhang et al. [27] proposed a driver speed model for curved roads
based on the recorded behavior of the driver.
Other works based their curve speed computation on the road shape estimation, either for
implementing it in semi-autonomous or autonomous solutions. Related to road geometry analysis,
identification of curved or straight segments has been performed either by analyzing vehicle’s ahead
motion [28–30] or by measuring physically-related parameters like curve radius, length and angle [6,7]
using GPS and GIS information. Once these curves are identified, their corresponding speeds are
calculated considering road friction and super-elevation angles [9,11,31].
Curve speed warning systems correspond to semi-autonomous solutions, which consist of
informing the driver, through sound or display, about the recommended speed [9,18,32,33], e.g.,
Varhelyi [18] proposed a system that goes beyond curves, considering also factors such as wet roads
and visibility conditions (darkness) to compute the appropriate speed and inform the driver visually
or audibly. Alternatively, the warning signal may come through some force applied to the throttle
(accelerator) to prevent speeding. That is the case of Huth et al. [34], who proposed a curve warning
system for motorcyclists through a force feedback throttle or a haptic glove.
Adaptive Cruise Control (ACC) systems, which are another type of semi-autonomous solution,
have also considered curve speeds together with speed limits to adjust velocity respecting headway
distance to the lead vehicle [10].
Concerning autonomous solutions, it is not a surprise that the automotive industry is trying
to develop control systems that consider every possible situation to drive safely. Nevertheless,
user acceptance is still a factor to overcome in order to bring autonomous vehicles to reality [18,21].
Through simulations, Park et al. [11] calculated curve speeds and applied them in a path-tracking algorithm.
Even though curve speed estimation has been considered in several works, only a few of
them have been actually implemented and tested either in simulations or real environments [9–11].
Glaser et al. [9], as well as Park et al. [11] have computed the ideal speed for curves based on the
analysis of geometric information, but no speed limits have been contemplated. On the other hand,
Lee et al. [10] considered speed limits and curve speeds, but their implementation is made for an
ACC, which is a semi-autonomous solution where the driver is able to take control of the vehicle and
override appropriate speeds. Our work is targeted to be implemented as an autonomous solution for
path-tracking algorithms.
where L is the distance between the front axle and rear axle (wheelbase) and α is the angle between the
vehicle heading vector and the look-ahead vector.
This algorithm suffers mainly from cutting corners (neglects curvature information) if look-ahead
distance is big and severe oscillations if it is defined as too small.
kε(t)
θ (t) = ψ(t) + arctan (2)
v(t)
This method is a simple basic approach that proved to be efficient in the DARPA race with
an average lateral offset of ±74 cm.
E
δ = arctan (1 − αS) (4)
R
where E is the distance between the axles of the car, R is radius of the circle drawn to follow the target
point and α is a coefficient set to 0.02.
Lombard’s method provides lateral errors smaller than 1 m when the speed is set to 36 km/h.
routes. We identified the noisy points and interpolated their new positions based on the previous and
following points.
Once GPS noisy points are removed, we computed the Euclidean distance between each pair of
points to keep GPS positions at an approximately 3.5-m distance between each other; this with the
aim to reduce the amount of data for further processing, i.e., road curvature estimation and speed
limits extraction.
We identified positions where speed limits change from 50 km/h to 30 km/h and vice versa.
By default, the speed limit at the starting point is set to 50 km/h if no other speed limit is identified.
An illustration of the procedure is presented in Figure 3.
The creation of the speed limits database is performed off-line and only after GPS information is
preprocessed. The description of the datasets and their speed limits identified is provided in Section 5.
Sensors 2017, 17, 1383 8 of 22
(
a)Re
pre
sen
tat
iono
f c
urv
e (
b)Be
ari
nga
ngl
ebe
twe
ens
egme
nts
p
ara
met
ers ABa
ndBC
Classifying points as being part of a curve or of a straight segment is the first step to be able to
identify start and end points (PT and PC) in a curve. This procedure is performed by computing the
bearing angle (α) between two consecutive segments formed with three points (A, B, C), as seen in
Figure 4b, using Equation (5).
!
−1 p (xB − x A )(xC − xB ) + (yB − y A )(yC − yB ) 180
α = cos p × (5)
2 2 2
( x B − x A ) + ( y B − y A ) × ( xC − x B ) + ( yC − y B )2 π
Once the bearing angle is computed, a threshold value is assigned to determine if the point B
is considered as part of a curve or not. If the angle α is bigger than the defined threshold, the center
point B is described as being part of a curve. Choosing the right threshold is a very important step,
since curvature identification depends on this value. In this work, a threshold value of 1.25◦ was
chosen as proposed in [6] after evaluating curvature identification with different values. It is worth
Sensors 2017, 17, 1383 9 of 22
mentioning that values smaller than 1.25 gave us false positives, as almost every point was considered
as being part of a curve, and values bigger than 1.25 discarded long and smooth curves.
Based on Li et al.’s work [6], curves were classified as simple or compound. A compound curve
is a curve formed with at least two consecutive curves separated by a certain distance between each
other. In our datasets, this distance was set to 10.5 m, which is the approximation of considering three
consecutive points in our preprocessed GPS path. An example of the two types of curves is illustrated
in Figure 5.
After each curve is defined by its start and end points (PT and PC), its geometric information,
consisting of its radius (R), length (L) and central angle (θ), is computed through Equations (6) to (8).
In order to perform these computations, the center point of the curve (O) is identified following the
work [6].
q
R = (xPC − xO )2 + (yPC − yO )2 (6)
q
C= (xPT − xPC )2 + (yPT − yPC )2 (7)
−1 C 180
θ = 2 × sin × (8)
2R π
Curve length (L) was estimated by summing up the Euclidean distances between the segments
that form the curve.
As we are interested in considering only dangerous curves, sharp curves were categorized if
their angle ranges from 30◦ to 180◦ , or their radius is between 5 and 18 m according to the American
Association of State Highway and Transportation Officials’ (AASHTO’s) “A Policy on Geometric
Design of Highways and Streets” [43]. Figure 6 shows an example of the curve classification according
to [6], and marked in red circles are the sharp curves identified with our method. In this work,
only sharp curves are considered to adjust the speed in path-tracking simulations.
Sensors 2017, 17, 1383 10 of 22
Figure 6. Curve classification of UTBM-2 dataset (see Section 5 for reference). Sharp curves are marked
with red circles.
Through the normal force (N), friction force ( f ) and weight force (mg) vectors, the centripetal
force is defined as:
mv2
FC = (9)
r
where m represents vehicle mass, v vehicle speed and r the curve radius.
Summing up the vertical components, normal force can be described following Equation (10).
mg
N= (11)
cos θ − µ sin θ
The net force is calculated through the horizontal components:
v2
(tan θ + µ)g = (14)
r
Using super-elevation e = tan θ and curvature information r = 1k , we substitute both terms into
Equation (14). The ideal speed will be given by Equation (15), which is exactly the same definition as
in [11].
r
(e + µ) g
v= (15)
k
Following this equation, convenient speeds for each sharp curve (vc ) were estimated considering
a super-elevation value between 6% and 12%, which, according to the American Association of State
Highway and Transportation Officials (AASHTO) [43], is the value defined for rural and urban roads.
k is the curvature information computed as the reciprocal of the radius, k = 1r .
Finally, given a sharp curve detected in the traveled GPS path, we computed the distance (dc ) from
the reference start point of the path to each curve starting point (PT) as shown in Figure 8. This distance
together with its respective curve lengths (lc ) and the convenient speed (vc ) are the parameters passed
to the speed negotiation algorithm, detailed in the next subsection.
Sensors 2017, 17, 1383 12 of 22
Figure 9. Curve driving scenario projecting vehicle position in the reference path.
Sensors 2017, 17, 1383 13 of 22
Once the projected vehicle position is known, the speed negotiation algorithm analyses which type
of road segment (curve or speed limits) is closer to the current vehicle position and computes a trigger
distance. This trigger distance (dtrig ) has the objective to take into account a smooth deceleration
behavior to reach the ideal speed with comfort. It is calculated with the following equation:
v2target − v2current
dtrig = (16)
2a
where:
• vtarget is either the speed limit (vl ) or curve speed (vc ),
• vcurrent is the current vehicle’s speed and
• a is the deceleration value to be applied.
An appropriate deceleration value is considered according to a study made by Maurya et al. [44],
which compares several deceleration behaviors at different speeds in various vehicle types. In our
case, we set a maximum acceleration (amax ) and deceleration (aneg ) of 2 m/s2 to provide a comfortable
transition between speeds and avoid longitudinal jerks.
If sharp curved positions overlap with zones where a speed limit is lower than 50 km/h,
our algorithm gives priority to the lower speed, which is usually the curve speed.
Considering all aspects discussed above, the speed negotiation algorithm works as follows:
At each time step, an ideal speed (videal ) results from the speed negotiation algorithm. This speed
is used by the lateral control system to calculate an appropriate angle (θ) to be applied by the vehicle
steering wheel. During every control cycle, the lateral control system sends to the vehicle speed and
angle parameters to provide a smooth tracking performance.
The goal of reducing the speed before arriving at a curve or a speed limit zone will provide the
passenger comfort, avoiding abrupt rapid decelerations. At the same time, respecting the ideal speed
definition will increase safety and provide more time in case hazardous situations occur. As a result,
lateral errors will decrease, which is the aim of path-tracking algorithms (shown in Algorithm 1).
5. Results
In this section, we perform simulations with different lateral control algorithms on a simulated
path and real datasets. This is with the aim to show how tracking accuracy is improved if correct speed
definitions are taken into account for each segment in the traveled path. Consequently, a sense of
comfort and safety will be provided to the passenger, reducing traffic accidents and severe injuries [1].
Three paths (one created and two real) were considered for the simulations. Each path has different
length, curve characteristics and speed limits. The real datasets (traveled paths) were acquired by
Qiao et al. [45] with the experimental platform shown in Figure 10 in Belfort, France. The data of the
paths are based only on GPS information captured by GPS and RTK-GPS receivers. The simulated
path was created with the aim to compare how noise affects the tracking performance of lateral
control algorithms.
A visual representation of the considered paths can be seen in Figure 11. They are briefly
presented below.
Figure 11. Datasets with sharp curves and speed limits marked. The normal black line represents a
speed limit of 50 km/h, the green line a speed limit of 30 km/h and the sharp curves in red.
1. Simulated path: This path is composed of 210 continuous points in a figure eight shape form.
Its total distance is about 374 m with four sharp curves identified. Since it is a virtual path, we set
the speed limit for the entire track to 50 km/h, as this is the default speed limit in urban areas in
France [24].
2. UTBM-2 dataset: The traveled path contains 541 points after the preprocessing step performed as
described in Section 4.1. The distance of this track is about 2.2 km with detected speed limits of
30 km/h in school zones. Six sharp curves were identified with different geometric characteristics.
Sensors 2017, 17, 1383 15 of 22
3. UTBM-3 dataset: This path is the longest one with a distance of about 4.5 km. It is represented by
1136 points after the preprocessing step. The number of sharp curves (dangerous curves) detected
was 17. The speed limits for this path are 30 km/h in residential areas and 50 km/h otherwise.
Different path tracking simulations were performed through the tool developed by
Lombard et al. [37]. These simulations include the analysis of different lateral control algorithms with
and without considering the ideal speeds obtained by our DSA method.
In order to compare the performance of our proposed DSA method between the different steering
control algorithms, (1) Pure Pursuit (PP), (2) Stanley, (3) Alice and (4) Lombard, we compute the root
mean square error (erms ) per each dataset for each of the four algorithms as follows:
s
∑( pd (t) − pc (t))2
erms = (17)
m
where pd and pc represent the desired and current vehicle position, respectively. The difference between
pd and pc is the lateral error, while m is the total number of sampling results every 400 ms (latency of
the system). Then, the average error on the total number of datasets per each method is calculated as:
1
erms =
N ∑ erms (18)
where N is the number of datasets per algorithm to be evaluated. In the case of curve evaluation,
N equals three (two real and one simulated datasets), while for speed limits, N equals two (the two
real datasets).
In France, the speed limit in urban areas is defined as 50 km/h [24], which is the maximum
speed assigned in the simulations. Other than that, in our datasets, speed limit zones of 30 km/h
were identified and considered to adjust vehicle’ speed. Regarding curve speeds, they vary from 14 to
36 km/h depending on the sharpness detected for each curve.
We analyzed lateral errors produced when speed limits change from 50 to 30 km/h in the UTBM-2
and UTBM-3 datasets (see Table 1). In other words, we analyzed lateral errors in segments of 30 km/h
(green zones as seen in Figure 11) and compare the results for each lateral control method. The speed
limit zone of 30 km/h in UTBM-2 is a straight segment, and even if lateral errors in all methods are
similar (15 to 20 cm error), the performance of all methods is improved considering our DSA method
(see Figure 12a). Lateral errors in the UTBM-3 dataset vary more due to its road geometry; therefore,
improvements with our DSA method are more noticeable, as shown in Figure 12b.
Table 1. Error comparison in zones of 30 km/h with different lateral control methods. Results are
expressed in meters. DSA, Dynamic Speed Adaptation; PP, Pure Pursuit.
Dataset Stanley Stanley DSA Alice Alice DSA PP PP DSA Lombard Lombard DSA
UTBM-2 0.0182 0.0158 0.0452 0.0416 0.1005 0.0554 0.0845 0.0550
UTBM-3 0.0210 0.0112 0.2840 0.1934 0.3346 0.2823 0.3374 0.2817
Average 0.0196 0.0135 0.1646 0.1175 0.2175 0.1688 0.2109 0.1683
On average, the evaluation of speed limits in segments of 30 km/h improved all of the methods
with our DSA, but not more than 5 cm, as seen in Figure 13. It is worth mentioning that the literature
has proven that respecting speed limits increases safety [2–4,19]. Autonomous vehicles capable of
adjusting automatically speed limits will be preferred, since, in real scenarios, drivers go beyond the
authorized speeds, being prone to cause accidents.
Sensors 2017, 17, 1383 16 of 22
Figure 12. Root mean square error (erms ) comparison of lateral errors obtained in 30 km/h speed limit segments.
Figure 13. Average root mean square error (erms ) of lateral error comparison between methods for
30 km/h speed limit segments.
Now, we will focus our analysis on the most dangerous segments, which are the sharp curves.
Tables 2–4 show lateral error results obtained by the different steering control algorithms in the detected
curves for each path.
Table 2. Error comparison in sharp curves of different Lateral Control methods for the simulated path.
Results are expressed in meters.
Curve Stanley Stanley DSA Alice Alice DSA PP PP DSA Lombard Lombard DSA
1 0.0599 0.0388 0.7637 0.6249 0.1935 0.2257 0.1854 0.0613
2 0.0762 0.0302 0.8286 0.6128 0.7972 0.0603 0.1752 0.04807
3 0.1029 0.0508 1.2195 0.9323 0.4351 0.2239 0.2223 0.09659
4 0.0493 0.0315 0.5757 0.4726 0.2107 0.1204 0.1531 0.04335
Average 0.0721 0.0378 0.8469 0.6607 0.4091 0.1576 0.1840 0.06233
Sensors 2017, 17, 1383 17 of 22
Table 3. Error comparison in sharp curves of different lateral control methods for the UTBM-2 dataset.
Results are expressed in meters.
Curve Stanley Stanley DSA Alice Alice DSA PP PP DSA Lombard Lombard DSA
1 0.0722 0.0288 0.5747 0.3308 0.8657 0.2451 0.4231 0.0552
2 0.0743 0.0245 0.6572 0.4945 0.7135 0.2372 0.5050 0.0589
3 0.2586 0.1797 1.2132 0.7815 1.5847 0.6940 0.9028 0.1292
4 0.0574 0.0265 0.5724 0.2591 0.5528 0.2396 0.2616 0.0382
5 0.1214 0.0449 1.0485 0.5650 1.0909 0.3627 0.5347 0.0751
6 0.1001 0.0441 0.9109 0.4528 1.4346 0.7075 0.7563 0.0687
Average 0.114 0.0581 0.8295 0.4807 1.0404 0.4143 0.5639 0.0709
Table 4. Error comparison in sharp curves of different lateral control methods for the UTBM-3 dataset.
Results are expressed in meters.
Curve Stanley Stanley DSA Alice Alice DSA PP PP DSA Lombard Lombard DSA
1 0.0763 0.0328 0.6984 0.3589 0.5012 0.2007 0.4840 0.0574
2 0.0910 0.0311 0.7711 0.4342 0.5934 0.1141 0.5959 0.0530
3 0.0545 0.0258 0.3409 0.2619 0.4462 0.1409 0.4498 0.0337
4 0.0421 0.0196 0.3409 0.1912 0.4056 0.0975 0.3857 0.0218
5 0.0893 0.0296 0.7401 0.4422 0.6162 0.0948 0.6225 0.0519
6 0.4570 0.4209 1.4465 1.0717 1.3936 1.3865 1.4099 0.2663
7 0.1697 0.1054 1.0596 0.6294 0.8554 0.565 0.8769 0.0835
8 0.1781 0.0705 1.1031 0.7269 0.9603 0.3007 0.9640 0.0818
9 0.0718 0.0297 0.3765 0.3331 1.1918 0.3802 1.2084 0.0523
10 0.0753 0.0278 0.5634 0.3694 1.0366 0.1566 1.0426 0.0488
11 0.0344 0.0183 0.2441 0.1671 0.5754 0.4024 0.5825 0.0240
12 0.0472 0.0247 0.4376 0.2123 0.2039 0.1248 0.2053 0.0306
13 0.0209 0.0193 0.2167 0.0885 0.1104 0.1312 0.1101 0.0145
14 0.1573 0.0669 0.9394 0.6268 0.9873 0.1431 0.9973 0.0613
15 0.1027 0.0389 0.9034 0.4698 0.5558 0.2250 0.5589 0.0651
16 0.1171 0.0424 0.7964 0.4465 0.9738 0.3490 0.8710 0.0706
17 0.1505 0.0615 0.5531 0.6594 0.934 0.2574 0.8212 0.0813
Average 0.1138 0.0627 0.6783 0.4405 0.7259 0.2982 0.7168 0.0646
We will start comparing the simulations performed between paths that contain or not noisy
information. Regarding tracking error results obtained in the simulated path (Table 2), we can see that
most of the lateral errors are the smallest (except for the Alice method) compared to errors obtained in
real datasets (Tables 3 and 4). A visual representation of this comparison can be seen in Figure 14a.
We can confirm that noise is an important factor for tracking algorithms, and it deserves special
attention to reduce it before controlling autonomous cars. In the literature, the authors have combined
different information to deal with noise, e.g., taking into account the dynamics of the car (position,
orientation, speed) [20].
Sensors 2017, 17, 1383 18 of 22
Figure 14. Root mean square error (erms ) comparison of lateral errors obtained in sharp curves for each dataset.
The improvements of each method, with and without considering speed regulations in curves,
range from 22% to 88% depending on the path. For example, in the simulated path (see Figure 14b),
tracking error is reduced by 19 cm in the Alice method (22% improvement). In UTBM-2, the method
with the largest root mean square error is pure pursuit (1.2 m), reduced by 72 cm with our DSA (see
Figure 14c). A visual representation of lateral errors on the UTBM-2 dataset with Pure Pursuit method
is shown in Figure 15, which illustrates how our DSA method is able to reduce errors significantly
in sharp curves. At the same time, it is clearly noticeable that the biggest lateral errors are present in
sharp curves confirming the need to adapt vehicle’ speed. Furthermore, in the UTBM-3 dataset (see
Figure 14d), tracking error is reduced by about 45 cm in the pure pursuit method, while for the Lombard
method, it is 69 cm (88% improvement).
In general, the performance of our DSA method in sharp curves showed significant (at least
10-cm error) improvements for most of the methods (see Figure 16), except for the Stanley method
(improvement of around 5 cm). The reason behind this behavior is the definition of its goal point. Since
the Stanley method is based on Pure Pursuit (PP) and the goal point is defined with a short distance,
it does not suffer greatly from the “cutting corners” effect.
Sensors 2017, 17, 1383 19 of 22
Figure 15. Speed (top graph) and lateral error (bottom graph) comparison in the UTBM-2 dataset using the PP
method with and without DSA.
Figure 16. Average root mean square error (erms ) of the lateral error comparison between methods for
sharp curves.
As tracking errors for speed limits decrease, but not significantly (1 cm for Stanley, 5 cm for Alice
and 3 cm for pure pursuit and Lombard), we will base our final conclusion on the results obtained
in sharp curves. Tracking errors with our DSA method improved three methods significantly, Pure
Pursuit (PP), Alice and Lombard. Even though the improvements for the Stanley method are not
considered significant, there is an improvement of about 5 cm when speed adaptation is used. The two
methods that benefit the most with DSA are pure pursuit and Lombard (as seen in Figure 16) with
about a 50- and a 47-cm difference respectively for sharp curves. For the Alice method, the difference
is about 30 cm, which is also a very good improvement. These results prove that, having the adapted
speeds to pass through sharp curves certainly leads to provide the passenger confidence and safety for
taking a ride in an autonomous vehicle.
One of the main limitations for our proposed method resides in GPS noise. Nevertheless,
this limitation can be treated combining more information, for example vehicle dynamics [20],
to analyze position, orientation and speed. Inaccurate curvature geometry extraction is another
limitation when noisy GPS information is present. Nonetheless, this last problem can be overcome
extracting curve parameters from digital maps.
Sensors 2017, 17, 1383 20 of 22
6. Conclusions
We have proposed a real-time Dynamic Speed Adaptation (DSA) method based on the analysis
of speed limits and curvature information. A speed limit database of the traveled paths was created
and considered as input for our DSA. Curvature information extraction allowed DSA to identify sharp
curves and to compute the recommended speed for each one. The speed limit database and curve
geometric information are extracted off-line from GPS paths in order to provide knowledge ahead of
the road while traveling in it.
The tracking performance using our approach showed significant reduction of lateral errors.
This improvement may potentially prevent accidents and reduce severe injuries in real driving
scenarios, which is exactly the aim of any autonomous vehicle.
The main advantage of our proposed DSA method is its adaptability, since it can be implemented
in the vehicle lateral control system independently of the steering wheel angle computation. Besides,
it enables the car to drive with recommended speeds through straight and curved segments respecting
traffic regulations. Furthermore, it is smart enough to localize upcoming sharp curves and apply a
comfortable deceleration value before arriving at them, reaching their ideal speeds to traverse them.
Future works will include the collection and analysis of vehicle dynamics in order to deal with
GPS noise. This information will correct noisy points in the path by comparing the vehicle’ speed and
orientation with the GPS data.
Furthermore, we will incorporate the information of a stereo system in order to bring perception
capabilities to the autonomous vehicle. The information of this sensor together with a semantic
segmentation approach will allow us to: (1) detect the road if GPS is not available; (2) identify speed
limit traffic signs in case no speed limit database is accessible; (3) recognize lane markings on the road
to guide the algorithm for road boundary detection and shape estimation; and (4) prevent the vehicle
from turning if the lane where it is driving has some curved arrows painted. Detecting and recognizing
lane markings on the road are a potential source of information for the autonomous car, since they
provide very useful knowledge to learn driving behaviors.
Acknowledgments: This research was supported by the National Council of Science and Technology (CONACYT).
Special thanks to Alexandre Lombard and Abdeljalil Abbas-Turki for providing us the code of the control
system simulator.
Author Contributions: Citlalli Gámez Serna conceived of the work, realized the algorithms, performed the
experiments, analyzed the data and wrote the manuscript. Yassine Ruichek provided many suggestions for the
work and made great efforts with respect to the paper revision.
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Vehicle Safety, Deliverable 4.8 u of the EC FP7 Project DaCoTA. 2012. Available online: http://www.
dacota-project.eu/ (accessed on 14 March 2017).
2. Carsten, O.M.; Tate, F. Intelligent speed adaptation: Accident savings and cost–benefit analysis.
Accid. Anal. Prev. 2005, 37, 407–416.
3. Driscoll, R.; Page, Y.; Lassarre, S.; Ehrlich, J. LAVIA–an evaluation of the potential safety benefits of the
French intelligent speed adaptation project. Annu. Proc. Assoc. Adv. Automot. Med. 2007, 51, 485–505.
4. Vlassenroot, S.; Broekx, S.; De Mol, J.; Panis, L.I.; Brijs, T.; Wets, G. Driving with intelligent speed adaptation:
Final results of the Belgian ISA-trial. Trans. Res. Part A Policy Pract. 2007, 41, 267–279.
5. Lai, F.; Carsten, O.; Tate, F. How much benefit does Intelligent Speed Adaptation deliver: An analysis of its
potential contribution to safety and environment. Accid. Anal. Prev. 2012, 48, 63–72.
6. Li, Z.; Chitturi, M.; Bill, A.; Noyce, D. Automated identification and extraction of horizontal curve information
from geographic information system roadway maps. Transp. Res. Rec. J. Transp. Res. Board 2012, 2291, 80–92.
7. Hans, Z.; Souleyrette, R.; Bogenreif, C. Horizontal Curve Identification and Evaluation; Technical Report;
Iowa State University: Ames, IA, USA, 2012.
Sensors 2017, 17, 1383 21 of 22
8. Neuman, T.R.; Glennon, J.C.; Saag, J.B. Accident Analyses for Highway Curves; Number 923; National Academy
of Sciences: Washington, DC, USA, 1983.
9. Glaser, S.; Nouveliere, L.; Lusetti, B. Speed limitation based on an advanced curve warning system.
In Pronceedings of the 2007 IEEE Intelligent Vehicles Symposium, Istanbul, Turkey, 13–15 June 2007;
pp. 686–691.
10. Lee, J.W.; Prabhuswamy, S. A Unified Framework of Adaptive Cruise Control for Speed Limit Follower and Curve
Speed Control Function; Technical Report; SAE Technical Paper: Detroit, MI, USA, 2013.
11. Park, M.; Lee, S.; Han, W. Development of steering control system for autonomous vehicle using
geometry-based path tracking algorithm. ETRI J. 2015, 37, 617–625.
12. Thrun, S.; Montemerlo, M.; Dahlkamp, H.; Stavens, D.; Aron, A.; Diebel, J.; Fong, P.; Gale, J.; Halpenny, M.;
Hoffmann, G.; et al. Stanley: The robot that won the DARPA Grand Challenge. J. Field Robot. 2006, 23, 661–692.
13. Alonso, I.P.; Llorca, D.F.; Gavilán, M.; Pardo, S.Á.; García-Garrido, M.Á.; Vlacic, L.; Sotelo, M.Á. Accurate
global localization using visual odometry and digital maps on urban environments. IEEE Trans. Intell.
Transp. Syst. 2012, 13, 1535–1545.
14. Gruyer, D.; Belaroussi, R.; Revilloud, M. Map-aided localization with lateral perception. In Proceedings of
the 2014 IEEE Intelligent Vehicles Symposium, Ypsilanti, MI, USA, 8–11 June 2014; pp. 674–680.
15. Welch, G.; Bishop, G. An Introduction to the Kalman Filter; University of North Carolina at Chapel Hill:
Chapel Hill, NC, USA, 1995.
16. Wang, S.; Deng, Z.; Yin, G. An Accurate GPS-IMU/DR Data Fusion Method for Driverless Car Based on
a Set of Predictive Models and Grid Constraints. Sensors 2016, 16, 280.
17. Wang, C.; Hu, Z.; Uchimura, K. Precise curvature estimation by cooperating with digital road map.
In Proceedings of the 2008 IEEE Intelligent Vehicles Symposium, Eindhoven, The Netherlands, 4–6 June 2008;
pp. 859–864.
18. Varhelyi, A. Dynamic speed adaptation in adverse conditions: A system proposal. IATSS Res. 2002, 26, 52–59.
19. Agerholm, N.; Waagepetersen, R.; Tradisauskas, N.; Harms, L.; Lahrmann, H. Preliminary results from the
Danish intelligent speed adaptation project pay as you speed. IET Intell. Transp. Syst. 2008, 2, 143–153.
20. Tradišauskas, N.; Juhl, J.; Lahrmann, H.; Jensen, C.S. Map matching for intelligent speed adaptation.
IET Intell. Transp. Syst. 2009, 3, 57–66.
21. Young, K.L.; Regan, M.A.; Triggs, T.J.; Jontof-Hutter, K.; Newstead, S. Intelligent speed adaptation—Effects
and acceptance by young inexperienced drivers. Accid. Anal. Prev. 2010, 42, 935–943.
22. Walker, G.H.; Stanton, N.A.; Young, M.S. The ironies of vehicle feedback in car design. Ergonomics 2006,
49, 161–179.
23. Spichkova, M.; Simic, M.; Schmidt, H.; Cheng, J.; Dong, X.; Gui, Y.; Liang, Y.; Ling, P.; Yin, Z. Formal models
for intelligent speed validation and adaptation. Procedia Comput. Sci. 2016, 96, 1609–1618.
24. French Government, Code de la Route. Article R413-1, R413-2, R413-3, R413-4. 2001. Available online:
https://www.legifrance.gouv.fr (accessed on 04 April 2017).
25. Cleret, J.; Chauvin, P.; Bernard, G.; Dupre, G.; Floris, O. Méthode de sélection des virages à signaler et niveau
de signalisation à implanter. In Guide Pratique, Conseil Général de la Seine-Maritime; DocPlayer.fr: Normandie,
France, 2001. (In French)
26. Lehtonen, E.; Lappi, O.; Koirikivi, I.; Summala, H. Effect of driving experience on anticipatory look-ahead
fixations in real curve driving. Accid. Anal. Prev. 2014, 70, 195–208.
27. Zhang, D.; Xiao, Q.; Wang, J.; Li, K. Driver curve speed model and its application to ACC speed control in
curved roads. Int. J. Automot. Technol. 2013, 14, 241–247.
28. Eidehall, A.; Pohl, J.; Gustafsson, F. Joint road geometry estimation and vehicle tracking. Control Eng. Pract.
2007, 15, 1484–1494.
29. Lundquist, C.; Schon, T.B. Road geometry estimation and vehicle tracking using a single track model.
In Proceedings of the 2008 IEEE Intelligent Vehicles Symposium, Eindhoven, The Netherlands, 4–6 June 2008;
pp. 144–149.
30. Lundquist, C.; Schön, T.B. Joint ego-motion and road geometry estimation. Inf. Fusion 2011, 12, 253–263.
31. Gámez Serna, C.; Lombard, A.; Ruichek, Y.; Abbas-Turki, A. Gps-Based Curve Estimation for an Adaptive
Pure Pursuit Algorithm. In Proceedings of the Mexican International Conference on Artificial Intelligence,
Cancún, Mexico, 23–29 October 2016; Springer: Berlin, Germany, 2016.
Sensors 2017, 17, 1383 22 of 22
32. Bonneson, J.; Pratt, M.; Miles, J. Evaluation of Alternative Procedures for Setting Curve Advisory Speed.
Transp. Res. Record J. Transp. Res. Board 2009, 2122, 9–16.
33. Chowdhury, M.A.; Warren, D.L.; Bissell, H.; Taori, S. Are the criteria for setting advisory speeds on curves
still relevant? Inst. Transp. Eng. ITE J. 1998, 68, 32.
34. Huth, V.; Biral, F.; Martín, Ó.; Lot, R. Comparison of two warning concepts of an intelligent Curve Warning
system for motorcyclists in a simulator study. Accid. Anal. Prev. 2012, 44, 118–125.
35. Snider, J.M. Automatic Steering Methods for Autonomous Automobile Path Tracking; Technical Report
CMU-RITR-09-08; Robotics Institute Carnegie Mellon University: Pittsburgh, PA, USA, 2009.
36. Linderoth, M.; Soltesz, K.; Murray, R.M. Nonlinear lateral control strategy for nonholonomic vehicles.
In Proceedings of the 2008 American Control Conference, Seattle, WA, USA, 11–13 June 2008; pp. 3219–3224.
37. Lombard, A.; Hao, X.; Abbas-Turki, A.; El Moudni, A.; Galland, S.; Yasar, A.-U.-H. Lateral Control of an
Unmaned Car Using GNSS Positionning in the Context of Connected Vehicles. Procedia Comput. Sci. 2016,
98, 148–155.
38. Szepe, T.; Assal, S.F. Pure Pursuit Trajectory Tracking Approach: Comparison and Experimental Validation.
Int. J. Robot. Autom. 2012, 27, 355.
39. Liu, R.; Duan, J. A path tracking algorithm of intelligent vehicle by preview strategy. In Proceedings of the
2013 32nd Chinese Control Conference (CCC), Xi’an, China, 26–28 July 2013; pp. 5630–5635.
40. Takasu, T.; Yasuda, A. Evaluation of RTK-GPS performance with low-cost single-frequency GPS receivers.
In Proceedings of the 2008 International Symposium on GPS/GNSS, Tokyo, Japan, 11–14 November 2008;
pp. 852–861.
41. Haklay, M.; Weber, P. OpenStreetMap: User-generated street maps. IEEE Pervasive Comput. 2008, 7, 12–18.
42. Kannan, S.; Thangavelu, A.; Kalivaradhan, R. An intelligent driver assistance system (i-das) for vehicle
safety modelling using ontology approach. Int. J. Ubicomp 2010, 1, 15–29.
43. AASHTO (American Association of State Highway and Transportation Officials). A Policy on Geometric
Design of Highways and Streets, 5th ed.; American Association of State Highway and Transportation Officials:
Washington, DC, USA, 2004.
44. Maurya, A.K.; Bokare, P.S. Study of deceleration behaviour of different vehicle types. Int. J. Traffic Transp. Eng.
2012, 2, 253–270.
45. Qiao, Y.; Cappelle, C.; Ruichek, Y. Place recognition based visual localization using LBP feature and SVM. In
Proceedings of the 14th Mexican International Conference on Artificial Intelligence, Cuernavaca, Mexico,
25–31 October 2015; Springer: Berlin, Germany, 2015; pp. 393–404.
c 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).