HAL Id: hal-03049723
https://hal.uca.fr/hal-03049723
Submitted on 10 Dec 2020
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Distributed under a Creative Commons Attribution| 4.0 International License
Map Matching and Lanes Number Estimation with
Openstreetmap
Abderrahim Kasmi, Dieumet Denis, Romuald Aufrère, Roland Chapuis
To cite this version:
Abderrahim Kasmi, Dieumet Denis, Romuald Aufrère, Roland Chapuis. Map Matching and Lanes
Number Estimation with Openstreetmap. 2018 21st International Conference on Intelligent Trans-
portation Systems (ITSC), Nov 2018, Maui, France. pp.2659-2664, �10.1109/ITSC.2018.8569840�.
�hal-03049723�
Map matching and lanes number estimation with OpenStreetMap
Abderrahim KASMI
1,2
, Dieumet DENIS
1
, Romuald AUFRERE
2
, Roland CHAPUIS
2
Abstract— Road information, like lanes number, play an im-
portant role for intelligent vehicles (IV). Traditionally such road
information are obtained through a vision-based measurement
or by using a digital detailed map. In this paper, we present
a new method for estimating the number of lanes using a low
precision GPS receiver and OpenSteetMap (OSM). The method
includes the integration of the GPS traces and OSM for a map
matching. To this end we developed a probabilistic multicriteria
algorithm for map matching that takes into account the
accuracy of the GPS data and the attribute information of
the road from OSM. Afterward we estimate the number of
lanes from OSM. We tested our algorithm on a set of GPS
data collected in an urban area near Paris for a total distance
of 50km and the overall estimation accuracy reached 83.64%
I. INTRODUCTION
In recent years, significant progress have been made in
Driver Assistance Systems (DAS) towards automated driv-
ing. Many applications, e.g., lane change assistance systems,
have been the subject of several research projects [1]. For
these applications knowing the number of lanes is crucial [2].
Therefore several methods have been proposed to identify
the lanes. In [3], a vision-based application is proposed to
estimate lanes number in road. The width of the road is
evaluated using an inverse perspective mapping (IPM), then
a Hough transform is applied to detect the lanes. In order
to identify the lanes, the Vanishing Points estimation (VPs)
based on Bayesian posterior probability are used in [4]. A
combination of a Statistical Hough Transform (SHT) [5]
with a Particle Filter (PF) is presented in [6] for lanes
detection and lanes tracking. These vision-based applications
show interesting results. However, situations like occlusions,
strongly differing illuminations, and unmarked or partly
marked lanes remain unresolved issues. Hence to override
these limitations, other researches work on using a precise
digital map to enhance the camera based systems for lanes
detection [7]. However one of the main drawback of these
accurate digital maps is the exorbitant cost to build them.
To overcome this constraint, researches have been made on
less-expensive solutions to estimate the number of lanes. A
number of researchers use GPS data to create a digital map in
order to extract road information such as road curvature and
lane number from it. In [8] an estimation of the number of
lanes was performed by collecting the GPS traces on a road
in order to calculate the Distribution Variance-Road Width
*This work has been sponsored by Sherpa Engineering and ANRT
(Conventions Industrielles de Formation par la Recherche).
1
Sherpa Engineering, R&D Department, La Garenne Colombes, France.
[a.kasmi, d.denis] @sherpa-eng.com
2
University of Clermont Auvergne, CNRS, SIGMA
Clermont, Institut Pascal, F-63000 Clermont-Ferrand, France.
FirstName.Lastname@uce.fr
Discrete Model (DV-RWDM) [9]. These methods depend
highly on the number of traces and only an accuracy of 60%
was achieved [8].
On the other hand, other researches work on using the
collaborative OSM mapping project for intelligent vehicles.
The OSM geodata includes rich information about road.
In addition to the spatial information, the geodata contains
detailed information such as the name, the limitation of
speed and the number of lanes [10]. As it is a collaborative
project by volunteers, the accuracy of the geospatial data
is undefined. In litterature, the use of the OpeenStreetMap
geodata is limited to navigation tasks such a localisation [11].
In order to determine the location of the vehicle on a
OSM ’link’ a Map-matching algorithm is performed. The
Map matching integrates positioning data with spatial road
network to identify the correct ’link’ on which the vehicle is
travelling. In [12] a survey on the map matching methods was
presented, the author classified the map matching algorithms
depending on the analysis procedure as follows:
• Geometric analysis: The most commonly used map-
matching algorithm is a simple search algorithm. In this
approach the closest ’node’ or ’shape point’ is matched
to the GPS position.
• Topological analysis: In this map matching algorithm,
the geometry as well as the connectivity of the links
are used. In [13] a map matching based on a weighting
system is proposed. Using the sensors measurement
about the vehcile position, vehicle speed, yaw angle,
a weighting score is calculated to choose the best link.
• Probabilistic analysis: This approach requires the def-
inition of an elliptical confidence region around the
position obtained from the GPS receiver [14] the error
region is then superimposed to the road network to
identify the correct link.
Other mathematical tools has been used to perfom the map
matching, in [15] a probabilistic criteria is calculated to select
the right link on which the vehicle travels. A Bayesian Belief
Network (BBN) is used, the network takes several inputs
like: the distance between the link and the vehicle position,
the difference angle, the accuracy of the GPS receiver, the
connectivity of the link and the direction of the link.
Compared with other related works, this paper presents a
new method for map matching and lane numbers estimation
exploiting OpenStreetMap. The map matching algorithm
proposed in this paper is a multicriteria algorithm that takes
into account several factors: geometrical, topological and
probabilistic. The three methods for computing the multi-
criteria depend on the availability of sensors on the vehicle.
For the number of lanes estimation, in contrast to [9], our
method does not require the collection of several GPS traces.
In addition, our algorithm was performed on a set of 6596
GPS points collected in an urban area for a total distance
of 50km. An accuracy rate of 82.95% on lanes number
estimation was reached.
The remainder of the paper is organized as follows: Section
II introduces the data model of the OSM geodata. Section
III describes the Map matching algorithm used. Section IV
presents the real-world experimental results. The paper will
be summarized and concluded in Section V.
II. THE OSM GEODATA
For local sections of the planet, an XML file containing
the latest revision of the OSM map can be downloaded
from the official website of OpenStreetMap. Furthermore,
regularly updates of the geodata are available at that website.
According to the OSM specifications, the OSM data model
consists of three basic geometric elements called Nodes,
Ways and Relations.
A. Nodes
Nodes n
i
are point-shaped geometric elements which are
used to represent GPS points in term of latitude (lat) and
longitude (lon). Moreover, Nodes are the basic points to
represent the geometry of the Way.
B. Ways
Ways are used to model line-shaped geometric objects like
roads, railways, pedestrianized road...etc. Throughout this
paper, a single Way is defined as follows:
W = (id
w
, N
w
, T
w
) (1)
Where id
w
denotes a unique identification number of the
Way. In addition, the set N
w
regroup the m Nodes n
i
representing the geometry of the Way as follows:
N
w
= {n
1
, n
2
, ..., n
m
} (2)
To specify the semantic of each Way, a subset T
w
of n tags
is related to the Way. According to the OSM specifications,
each Way can have up to 255 tags. These tags are defined
as follows:
T
n
= {t
1
, t
2
, ..., t
n
} (3)
where t
i
indicates a single tag. Each tag consists of two
elements, a key k and the corresponding value v :
t
i
= (k, v) (4)
As illustrated on Figure 1 an example of an OSM Way with
its corresponding tags.
C. Relations
The element relation describes the relationship between
Nodes as well as Ways. This element is not used in this
paper. For more information, refer to [16].
.
Fig. 1: Example of an OSM Way with its corresponding tags,
keys and values.
TABLE I: Used OSM tags
Description Key Value
road speed limit in km/h maxspeed number
total number of physical lanes of the Way lanes number
D. The proceeding of the OSM data
For this paper, only two OSM tags from the 255 possible
are used. The tags used are listed in the Table I.
The OSM map consists of a set of Ways. To increase
the robustness of our map matching a preprocessing stage
is performed. Only Ways representing roads on which the
vehicle may travel are selected as shown on Figure 2.
III. MAP MATCHING ALGORITHM
As discussed before, an OSM map consists of a set of
Ways, each Way is constructed from multiple Nodes. In our
work we represent each Way by a set of segments. In other
words, belonging to segment is equivalent to belonging to an
OSM Way. So the map matching task can be reformulated
as matching a GPS point with a segment. In the remaining
sections, words segment and Way are switchable.
To identify the Way matched with a GPS point, it is necessary
to perform for each GPS node a ’discrimination stage’ to
remove all the Ways which are not suitable for the map
matching. Afterwards, an OSM Way is chosen from the
remaining OSM Way candidates by calculating a probabilis-
tic criterion. In our paper, three methods for calculating
(a) Before (b) After
Fig. 2: OSM map before preprocessing stage Fig (a) and
after preprocessing stage Fig (b)
this criterion are presented. Finally a Way with unique
identification number is selected. Using this identification
number, the number of lanes is extracted from the OSM
geodata. A general presentation of the algorithm is shown in
Figure 3.
In the following discussions, we will first illustrate the main
factors we adopted for the discrimination stage. Thereafter
we present the probabilistic criterion adopted to the map
matching.
Fig. 3: Presentation of the algorithm
A. Discrimination stage
The aim of this stage is to remove all the Ways which are
not suitable for map matching depending on several factors:
1) Distance to road: The distance between two geograph-
ical objects can be represented by three types of distances. As
illustrated in Figure 4, this distance can be a perpendicular
distance (d
0
) or a distance from point to an end point (d
1
and d
2
).
In our work, we take the closest distance as being the shortest
distance to a Way (d
0
in Figure 4a and d
2
in Figure 4b).
Fig. 4: Closest distance between a GPS position and a Way
(d
0
in (a) and d
2
in (b))
Ways with a distance smaller than a threshold are picked
as Way candidates.
2) Angle difference: As shown in Figure 5, we compute
the angle difference between the GPS trace and the Way. We
retain the Ways that have an angle difference less than 90
◦
.
3) Speed difference: In addition to lon, lat information,
the GPS receiver provides information about the speed of
the vehicle, we compare this speed to the value of the tag
Fig. 5: Difference of angles α
1
, α
2
between GPS trace and
Ways W
1
, W
2
”maxspeed” for every Way candidates. If the speed of the
vehicle is greater than a threshold then the Way is eliminated.
The threshold is defined as the sum of the limitation speed,
from the tag ”maxspeed”, plus 40km/h.
B. The probabilistic criteria
The distance and orientation criteria are not always suf-
ficient to select the right Way. There might still be an
ambiguity on choosing the right Way, as pictured on Figure 6.
Fig. 6: Ambiguity to select the correct Way at t
k
between
W
1
, W
2
and W
3
In order to choose the right Way, we define a probabilistic
selection criterion allowing to take into account the uncer-
tainty on the choice of the Way on which the vehicle travels.
In this paper, three methods of calculating this probabilistic
criterion are presented:
• Probabilistic criterion based on Euclidean distance,
• Probabilistic criterion based on Mahalanobis distance,
• Probabilistic criterion based on the probability of be-
longing to a segment.
1) Probabilistic criterion based on Euclidean distance:
The map matching task can be formulated as the calculation
of the highest posterior probability of a GPS measurement
Z
k
belonging to a Way W
i
:
arg max
i
P (W
i
|Z
k
) (5)
By assuming the uniformly distributed prior probability for
the Ways P (W
i
) and using Bayes formula [17], we get:
arg max
i
p(W
i
|Z
k
) = arg max
i
(
p(Z
k
|W
i
)P (W
i
)
p(Z
k
)
) (6)
Since p(Z
k
) is constant for each Way candidate, then we
have:
arg max
i
p(W
i
|Z
k
) ≡ arg max
i
(p(Z
k
|W
i
)P (W
i
)) (7)
Since p(W
i
) is uniformly distributed, we get:
arg max
i
P (W
i
|Z
k
) ≡ arg max
i
p(Z
k
|W
i
) (8)
The probability p(Z
k
|W
i
) is modeled by two random
independent variables :
• The distance to a road d is modeled as zero mean,
normally distributed random variable with standard de-
viation σ
d
described as follows:
σ
d
= .DOP
1
(9)
Due to the uncertainty of the GPS data and the OSM
map (beetween 6-9 m [16]). The theoretical error in
distance was chosen to be 15 m.
• The angle difference θ is modeled as zero mean, nor-
mally distributed random variable with standard devia-
tion σ
θ
defined as being inversely proportional to the
speed of the vehicle v:
σ
θ
∝
1
v
(10)
We rewrite (8) as follows:
arg max
i
P (W
i
|Z
k
) = arg max
i
(p
d
(d(Z
k
, W
i
)) ×p
θ
(θ(Z
k
, W
i
))
= arg max
i
(
1
√
2π × σ
d
.exp(−
d
2
(Z
k
, W
i
)
2σ
2
d
).
1
√
2π × σ
θ
.exp(−
θ
2
(Z
k
, W
i
)
2σ
2
θ
))
= arg min
i
d
2
(Z
k
, W
i
)
σ
2
d
+
θ
2
(Z
k
, W
i
)
σ
2
θ
(11)
Finally, the probabilistic criterion C
d
is calculated for every
way candidate, the Way having the smallest criterion is
chosen among the Way candidates:
C
d
=
d
2
(Z
k
, W
i
)
σ
2
d
+
θ
2
(Z
k
, W
i
)
σ
2
θ
(12)
This selection criterion does not take into account the cor-
relation between the GPS measurements in longitudinal and
lateral directions. In addition, the orientation angle of the
vehicle is calculated from the GPS trace. To address these
issues, an extended Kalman-filter (EKF) is used in order to
get the uncertainty matrix related to the pose of the vehicle.
2) Probabilistic criterion based on Mahalanobis distance:
Using the covariance matrix associated with the pose of the
vehicle Σ
X
, we compute the Mahalanobis distance for each
Way candidate. To this end, we calculate the orthogonal
projections of the vehicle position onto every Way candidate.
Afterward, the Mahalanobis distance C
m
is calculated as
follows:
C
m
=
q
(X − Y )
T
Σ
−1
X
(X − Y ) (13)
• X = (x, y, θ)
T
the current state vector representing the
pose of the vehicle and X its estimation and Σ
X
the
corresponding covariance matrix,
• Y = (x
w
, y
w
, ψ)
T
the vector representing the orthogo-
nal projection on the Way.
1
Dilution of precision
3) Probabilistic criterion based on the probability of be-
longing to a segment: To compute this probabilistic criterion,
we look for the probability of a segment [AB] defined by
A = (x
A
, y
A
)
T
and B = (x
B
, y
B
)
T
to belong to an expected
area of presence for the vehicle.
To this end, each segment [AB] is modeled as a random
vector S
AB
centered on S
AB
and defined by its covariance
matrix Σ
AB
as follows:
S
AB
∼ N (S
AB
, Σ
AB
) (14)
S
AB
=
(x
B
+ x
A
)/2
(y
B
+ y
A
)/2
(15)
Σ
AB
= V LV
−1
(16)
V =
x
B
− x
A
y
A
− y
B
y
B
− y
A
x
B
− x
A
(17)
L =
λ
1
0
0 λ
2
=
|AB|
2
4
0
0
e
2
s
4
!
(18)
With e
s
being the thickness of every link, λ
1
and λ
2
the
eigenvalues of the the matrix V . Thus the probability we are
looking for is defined as follows:
C
p
= P
AB
=
Z
N(
¯
X, Σ
X
)N(S
AB
, Σ
AB
)dX (19)
The product of two Gaussian distributions is a denormalized
Gaussian distribution, we can rewrite (19) as follows:
C
p
=
R
N (
¯
X, Σ
X
)N (S
AB
, Σ
AB
)dX =
R
kN(µ, Σ)dX (20)
Using the canonical representation for a Gaussian [18], we
get:
C
p
= k =
1
p
exp(g
1
+ g
2
+
1
2
µ
T
Σ
−1
µ) (21)
with:
Σ = [Σ
−1
X
+ Σ
−1
AB
]
−1
µ = [Σ
−1
X
+ Σ
−1
AB
]
−1
[Σ
−1
X
¯
X + Σ
−1
AB
¯
S
AB
]
p = log [(2π)
−n/2
|Σ|
−1/2
]
g
1
= log [(2π)
n/2
|Σ
X
|
−1/2
] −
1
2
¯
X
T
Σ
−1
X
¯
X
g
2
= log [(2π)
n/2
|Σ
AB
|
−1/2
] −
1
2
¯
S
T
AB
Σ
−1
AB
¯
S
AB
C. Navigation history
The criteria calculated above do not take into account the
relationship between two successive position measurements.
Another criterion defining this relationship is introduced.
This one is based on a weighting concept, if a Way is
candidate at two consecutive time measurement intervals
(t
k−1
and t
k
), then this Way is more likely to be the selected
as the correct one. Figure 7 illustrates this case, the Way (W
1
)
is more likely to be the correct Way than the Way (W
3
) at
time t
k
.
We update the previous criteria calculated on Section III-B
by adding a weight w as follows:
Cn
i
= C
i
× w, i ∈ [d, m, p] (22)