[ieee 2010 13th international ieee conference on intelligent transportation systems - (itsc 2010) -...
TRANSCRIPT
Using Recursive Bayesian Estimation for Matching GPS Measurements
to Imperfect Road Network Data
Oleksiy Mazhelis
Abstract— Map-matching refers to the process of projectingpositioning measurements to a location on a digital road net-work map. It is an important element of intelligent transporta-tion systems (ITS) focusing on driver assistance applications,on emergency and incident management, arterial and freewaymanagement, and other applications. This paper addresses theproblem of map-matching in the applications characterized byimperfect map quality and restricted computational resources- e.g. in the context of community-based ITS applications.Whereas a number of map-matching methods are available,often these methods rely on topological analysis, thereby makingthem sensitive to the map inaccuracies. In the paper, a newmap-matching method based on the probabilistic approach isintroduced. In the method, the probabilities of alternative roadlinks are estimated with recursive Bayesian estimation, and theroad link is identified using maximum a posteriori probabilityprinciple. The topological analysis is not used; instead, thedistance between projections on road links is used to assignroad link switching probability. The accuracy of the methodis empirically evaluated, and the link identification accuracy isfound to be similar to that of alternative approaches relying ontopological analysis.
I. INTRODUCTION
The functionality of intelligent transportation systems
(ITS) often includes the so-called map-matching, i.e. the
process of finding the location of an object on a road,
whereby the positioning information is fused with a (digital)
road network map [1]. The process of map-matching usually
consists of identifying the road link and defining the position
of the object on this link. The result of map-matching can
be used both to determine physical location of the object,
and to improve the accuracy of positioning by correcting
positioning errors.
The correctness of the map-matching greatly depends on
the accuracy of positioning and the quality of the map
used [2]. Positioning may rely on a global navigation satellite
system such as Global Positioning System (GPS) or Galileo,
inertial navigation systems (INS) based on accelerometers
and gyro, or on another navigation approach, such as radio
and radar navigation. Nowadays, GPS receivers are available
as inexpensive consumer-grade electronic devices, including
the receivers embedded in numerous smartphone models.
The accuracy of these consumer-grade receivers yields to
the accuracy of high-end devices, and may be as low as over
70m for the first fix, eventually improving to circa 20 m [3].
This research reported in this paper was carried out in Sensor DataFusion and Applications project as a part of the Cooperative Traffic researchprogram of the Strategic Center for Science, Technology and Innovation inthe Field of ICT, funded by the National Technology Agency of Finland.
O. Mazhelis is with the Department of Computer Science and InformationSystems, University of Jyvaskyla, P.O. Box35, FIN-40014, Jyvaskyla,Finland (email: [email protected])
Road maps often also contain omissions and inaccuracies
in a form of road curves interpolated as linear segments,
missing links, missing information about maintenance works
or new roads [2], [4], [5]. These inaccuracies may make the
map-mapping based on topology problematic. Besides, an
implicit assumption often exists that the object stays on the
road, therefore excluding off-road motions altogether.
Map-matching is often used in cooperative ITS applica-
tions; an example of such application is pothole detection
based on the acceleration measurements from a fleet of ve-
hicles [6]. These applications may rely on open community-
based map providers such as OpenStreetMap and utilize the
GPS receivers embedded into Nokia N95 phones [7]. In this
case, the task of map-matching is challenging both due to the
positioning inaccuracies of consumer-grade GPS receivers,
and due to the inaccuracy of the map used.
This paper focuses on the problem of map-matching in the
application domains characterized by imperfect map quality
and restricted computational resources, such as the coop-
erative ITS applications exemplified above. It is therefore
assumed that a relatively cheap GPS receivers are used, and
the road network map is assumed to contain omissions and
inaccuracies. The matching is assumed to be done on-line,
and therefore, post-processing is to be avoided. Furthermore,
it is implied that the matching software will be deployed
as a part of a distributed solution, where the concurring
requests from multiple traffic participants need to be served;
in such environment, the computational overhead of the map-
matching method should be conservative to make the solution
scalable.
Whereas a number of map-matching methods are avail-
able, often these methods rely on topological analysis,
thereby making them sensitive to the map inaccuracies.
In this paper, a new map-matching method based on the
probabilistic approach is introduced. In the method, the
probabilities of alternative road links are estimated with
recursive Bayesian estimation, and the correct road link is
identified using maximum a posteriori probability principle.
The topological analysis is not used; instead, the distance
between projections on road links is used to assign road
link switching probability. The accuracy of the method is
empirically evaluated; according to the results, the estimated
probability of correctly identifying road links is similar to
that of alternative approaches relying on topological analysis.
The remainder of the paper is organized as follows.
In the next Section, the related map-matching approaches
are reviewed. In Section III, the map-matching method
based on recursive Bayesian estimation is described. The
2010 13th International IEEEAnnual Conference on Intelligent Transportation SystemsMadeira Island, Portugal, September 19-22, 2010
TC8.3
978-1-4244-7659-6/10/$26.00 ©2010 IEEE 1492
results of performance evaluation experiments are reported
in Section IV. Finally, in Section V, the advantages and
shortcomings of the method are discussed, and the directions
for future work are outlined.
II. RELATED WORKS
A number of map-matching approaches and algorithms
have been developed over last decade varying from geometric
and topological analysis to probabilistic and other advanced
approaches [1]. The geometric approaches rely on searching
for the map node closest to the position fix (point-to-point
matching), on searching the map curve closest to the fix
(point-to-curve matching), or on assessing the similarity
between measured trajectory and the road network (curve-
to-curve matching) [8]. These methods are rather sensitive
to outliers in the positioning measurements, and they were
also found to produce unstable results in dense urban envi-
ronments [1].
The topological approaches, in addition to the geometric
analysis, take into account the connectivity among the road
links [9], [10]. Probabilistic approaches construct an error
region around position measurement, and the likely road link
in this region is identified based on geometry and topological
information [2], [11]. Bayesian inference [12] and particle
filters [13] can be employed in order to estimate the road
link probability.
Besides probabilistic reasoning, the likelihood of road
links can be assessed using Dempster-Shafer evidential
reasoning [14], [5] or fuzzy logic [15]. Furthermore, the
map-matching process is often augmented with (Extended)
Kalman Filter, whereby the fused positioning and motion-
related information from multiple sensors is corrected by
using the map matching results [16], [17], [18], [5].
Majority of map-matching methods, in addition to the po-
sitioning information, also take into account the information
about the heading of the object, e.g. inferred from changes
in gyroscope readings [18] or from sequential positioning
measurements [5]. This information, derived from GPS mea-
surements, is utilized in the proposed method as well.
In many methods, the matching does not take into account
the preceding measurements. Yang et al. [14] take into
account the distance to the closest road at the present point
in time only. Similarly, Greenfeld et al. [9] and Velaga et
al. [10], when computing the total weighted score for the
heading, proximity, connectivity, and other measurements,
ignore the prior measurements. This reliance on the present
measurements makes the methods sensitive to the outliers.
In the proposed method, in order to benefit from the ac-
cumulated information, the recursive Bayesian inference is
applied.
Noteworthy, the road link connectivity information is used
in many approaches. For instance, the multiple hypothesis
technique [12], the “probability of connectivity” is estimated
based on the number of bypassing links between prior
and current mapped positions. Such connectivity analysis,
however, would make the method sensitive to the omissions
and inaccuracies of the digital road maps. Besides, it would
increase the computational complexity of the method, as
the number of bypassing links between prior and current
projections needs to be retrieved for each pair of projections.
Therefore, the analysis of road link connectivity is avoided
in the proposed method.
Arguably, the method closest to the proposed one is
the multiple hypothesis technique (MHT) [12]. Similarly
to the MHT, a set of candidate road links are identified,
and the candidate probabilities are estimated recursively.
However, to tolerate road map omissions and inaccuracies,
in the proposed method, instead of the road link connectivity
analysis, the link change probability is estimated based on
the distance between the position fix projections.
III. MAP MATCHING USING RECURSIVE BAYESIAN
ESTIMATION
We assume that a sequence of positioning measurements
are available, and the measurement vector at time t takes the
form zt = {PN, PE}, where PN and PE are north and east
positions of the object, respectively. The map-matching prob-
lem is formulated in the paper as follows: given a sequence
of positioning (GPS) measurements z0:t = {z0, z1, . . . , zt},
accumulated by time t, find the most likely road link st and
location on the links xt.
In the method, the probabilistic Bayesian theory is applied.
Namely, the a posteriori probability of road links P (st|z0:t)is approximated, and the maximum a posteriori probability
(MAP) principle is used to determine the appropriate road
link. The process of map-matching is recursive; on each
iteration, a new measurement zt is used to update P (st|z0:t).A single iteration consists of i) hypothesis generation, ii) hy-
pothesis evaluation, and iii) hypothesis selection phases,
which are described in detail below.
Hypothesis generation. Given a measurement, a set
of candidate road links is generated. For the initial map-
matching, an error ellipse is constructed around the initial
GPS measurement, and the candidate road links are selected
as the links crossing the error ellipse.
For each consequent measurement, first, the set of candi-
date links from the preceding step is extended with the road
links crossing the error ellipse constructed around the new
measurement. After that, the links in the set, whose distance
to the latest measurement exceeds a pre-defined threshold,
are pruned.
The candidate road links are employed to construct the
set of Nt hypotheses St ≡ {Sit, i = 1, . . . , Nt}. Individual
hypothesis Sit is defined as a vector Si
t = {sit,y
it, θ
it}, where
sit is the hypothesized road link, y
it is the projection of zt
on sit, and θi
t is the heading of the road link at the point of
projection.
Hypothesis evaluation. From Bayes rule, the conditional
probability P (sit|z0:t) can be expressed as:
P (sit|z0:t) ∝ P (z0:t, s
it) = P (zt|s
it)P (si
t|z0:t−1). (1)
Using the complete probability formula, the second term
1493
in the right-hand side of Eq. 1 can be rewritten as:
P (sit|z0:t−1) =
∑
sj
t−1∈St−1
P (sit|s
jt−1)P (sj
t−1|z0:t−1). (2)
Thus,
P (sit|z0:t) =
1
CP (zt|s
it)
∑
sj
t−1∈St−1
P (sit|s
jt−1)P (sj
t−1|z0:t−1),
(3)
where C represents the normalization coefficient.
The term P (sjt−1|z0:t−1) represents a recursive element.
The other terms in the right-hand side of Eq. 3 are estimated
as follows.
The probability of a measurement given a road link
P (zt|sit) is assumed inversely proportional to the distance
and the heading discrepancy, namely:
P (zt|sit) = Pproximity(zt|s
it)Pheading(zt|s
it). (4)
The proximity probability Pproximity(zt, sit) is inversely
proportional to the distance between zt and sit:
Pproximity(zt|sit) =
1
d(zt,yit) + δ
, (5)
where d(zt,yit) is the distance between zt and y
it, and δ
represents a small positive real number to limit the proximity
probability value.
Similarly, the heading probability Pheading(zt, sit) is
inversely proportional to the discrepancy between the
measurements-derived heading and the road link heading θit:
Pheading(zt|sit) = 1 − 2k
h(θ, ψ)
π, (6)
where ψ is bearing from zt−1 to zt estimated using the flat
earth approximation:
ψ = arctancosPNt−1(PEt − PEt−1)
(PNt − PNt−1)(mod 2π), (7)
h(θ, ψ) ≤ π/2 is the heading deviation between θ and ψ,
and 0 < k ≤ 1 is a regularization coefficient specifying
how sensitive the estimation of the heading discrepancy
probability is to the heading deviation. Note that the heading
deviation is calculated as the smallest angle between the lines
going in the directions θ and ψ. Thus, the heading deviation
ignores the information about one-way directions.
As mentioned above, the link switching probability esti-
mation usually takes into account the road network topology,
e.g. in terms of the road links bypassed [12]. However, this is
sometimes problematic: i) if preceding road link is wrongly
estimated and is not connected to the correct one, the proba-
bility will be close to zero; ii) it will take extra computational
time to retrieve and process the road connectivity from road
network map database; iii) in case a road link connecting the
preceding and the current links is missing in the road network
map, the connectivity information will contain a significant
error.
Therefore, the road link switching probability P (sit|s
jt−1)
is assumed inversely proportional to the square of the dis-
tance from the projection (of measurement zt−1 at time t−1)
onto road link sjt−1 to the projection (of measurement zt at
time t) onto road link sit:
P (sit|s
jt−1) =
{
ci1
[d(yit−1
,yjt)+ξi]2
, if i 6= j
ci1
[d(zt−1
,zt)]2 , if i = j.
(8)
where ci are normalization coefficients. Note that the no-link-
switching probability (i.e. the case of i = j) is estimated
based on the object’s measured displacement d(zt−1, zt).That is, as soon as the vehicle approaches a crossroad (see
Fig. 1), the distance between projections on the crossing road
links, as compared to the displacement, decreases, thereby
increasing the road link switching probability.
z0
z1
i0y
sj
z2
i1y
j2y
j1y
!jid 10 ,yy
!10 ,zzd
!jid 21,yy
!10 ,zzd
si
Fig. 1. The distances between projections d(yit−1
,yjt ) and measurement
displacements d(zt−1, zt). Their values are utilized to estimate the road
link switching probability P (sit|s
jt−1
): closer to a crossroad, the distancebetween projections on the crossing road links, as compared to the displace-ment, decreases, thereby increasing the road link switching probability
The smaller the distance between projections on two
road links, the higher the probability of switching the
links. This may result in erroneous switching between two
parallel roads. Therefore, the auxiliary non-negative term
ξi = d(zt−1, zt) is used in Eq. 8: it assigns greater values
to no-link-switching probability, and therefore increases the
threshold of switching the road links.
Hypothesis selection. The most likely road link is
identified by following MAP principle:
st = arg maxsi
t
P (sit|z0:t), (9)
where P (sit|z0:t) is estimated based on Eq. 3 for all hypothe-
ses in St.
The position of the object on the road link is assumed to
be xt = yt, where yt is the point on road link st closest
to zt. Such simple position determination does not take
into account the dynamics of object movements; in order to
improve the accuracy, the positioning can be smoothed using
Kalman filter (applied to measurements zt or to mapped
positions xt).
1494
At present, the method always outputs a decision (i.e. st
and xt, unless no road links in the available network map
cross the area used for hypothesis generation. In order to
improve the mapping accuracy, ambiguous decisions can be
filtered by imposing a minimum probability threshold and/or
a minimum ratio between the two largest probabilities; in
case the thresholds are not exceeded, a “no decision” would
be produced as an output.
IV. PERFORMANCE EVALUATION
In order to empirically evaluate the performance of the
proposed map-matching method, it has been applied to sev-
eral data sets. The datasets and the results of experimentation
are described below.
The positioning data used in the experiments was gathered
by using a Nokia N95 phone, which receives GPS readings
from embedded GPS receiver at the frequency of 1 Hz. The
data were stored on the phone’s memory card, and were then
processed offline.
The data were collected during summer and autumn 2009
in the course of 18 drives, all of which took place in the
region of Central Finland. The description of the gathered
datasets is provided in Table I. As could be seen, both
urban, suburban, and rural areas are represented. The urban
characteristics include buildings (though, very tall buildings
were not present due to the specifics of the area), dense
road networks, bridges and flyovers. The length of the drives
varied from 1.5 km for the shortest urban drive to over 70 km
for the longest (largely rural) drive.
In order to test the accuracy of the proposed map-mapping
method, the OpenStreetMap1 was used as a digital road
network map. Since this is a community project, the map
relies on users uploading their GPS track logs and editing the
vector data; as a result, the map is likely to have omissions
and inaccuracies.
1http://www.openstreetmap.org/
TABLE I
POSITIONING DATASETS
Dataset Environment Data points
1 Urban 3192 Urban 1643 Urban 1404 Urban 1595 Urban and suburban 4256 Urban and suburban 4337 Urban and suburban 2798 Urban and suburban 1209 Urban and suburban 42110 Urban and suburban 57911 Urban and suburban 46012 Urban and suburban 45513 Urban and suburban 43014 Urban and suburban 44315 Urban, suburban and rural 294716 Urban, suburban and rural 316317 Urban, suburban and rural 249618 Suburban and rural 2225
The software implementing the proposed map-matching
method is written in Java and is utilizing LibOsm library2.
For practical reasons, the error square bound was employed
instead of error ellipse in hypothesis generation phase; the
square with side of 50 m was used. Whenever a break in
measurements exceeded the threshold of 20 s, the measure-
ments were treated as discontinued and the map-matcher was
reinitialized. The map-matching was carried out in a batch
mode, and the results were processed off-line.
An example of the map-matching result is shown in Fig. 2.
As could be seen, the GPS positions are correctly mapped
onto the appropriate road links. The mapping accuracy was
evaluated by manually comparing the results of matching
with the known routes taken in the data gathering drives.
Only the accuracy of link identification was assessed; the
evaluation of the horizontal accuracy was not possible, since
the true position on the links was not known. Given position
mapped on a link, the mapping was considered correct, if
the corresponding link was driven on and if the order of the
links matched with the sequence of the link within the route
taken; otherwise, the mapping was considered erroneous.
Fig. 2. Example: raw measurements (red circles) and their mappedcounterparts (blue triangles)
It should be noted that the current method implementation
does not take into account the existence of parking lots,
neither does it distinct the opposite lanes. Therefore, neither
the erroneous mapping onto the neighboring road links while
maneuvering on a parking lot nor the mapping onto the
opposite lane were considered as a mapping error.
The results of accuracy evaluation are provided in Table II,
where both the correctly identified links and the link identifi-
cation errors are reported. As can be seen from the table, for
urban datasets, the link identification accuracy varies in the
2http://sourceforge.net/apps/mediawiki/
travelingsales/index.php?title=LibOSM
1495
range 95.7–100%, and for the mixed datasets in the range
87.5–99.9%. The link identification accuracy of the method is
therefore comparable to the results achieved with alternative
approaches utilizing topological data: as observed in [1], the
accuracy achieved with topological approaches varies from
85.8% [19] (suburban) to 99.2% [15] (urban and suburban).
Meanwhile, the proposed method is expected to have some
advantage as compared with the approaches relying on
topology. Namely, as link connectivity is not taken into
account, omissions and inaccuracies in the digital road map
are tolerated. Besides, the topological analysis, if used, would
add to the computational overhead of the map-matching
method, and therefore would make the implementation of
the map-matching software less scalable.
For urban and mixed environments, the obtained link
identification accuracy is summarized in Table III, and Fig. 3
shows the link identification errors while categorizing them
according to the environment. As could be seen, for the
mixed datasets with prevailing suburban and rural data,
the link identification accuracy is slightly better. The more
accurate mapping in suburban and rural areas can be ex-
plained by the scarcity of buildings and less dense and
simpler road network in these environments. It should be
noted that, as shown in the figure, for one of the datasets
representing mixed suburban and urban data (dataset 7), the
link identification errors exceed 12%, which is significantly
worse than the errors observed with the urban datasets. A
further inspection of the dataset has revealed that such an
exceptionally large percentage of errors was caused by a bias
in the positioning measurements.
V. DISCUSSION AND CONCLUDING REMARKS
In the previous sections, the map-matching approach based
on recursive Bayesian estimation was described, and the re-
sults of empirical evaluation of the approach were presented.
The achieved link identification accuracy is comparable to
the accuracy of alternative map-matching approaches. Since
TABLE II
ROAD LINK IDENTIFICATION ERRORS
Dataset Link Link Correctly Correctidentification identification identified link
errors errors (%) links identification (%)
1 0 0.0 319 100.02 6 3.7 158 96.33 6 4.3 134 95.74 2 1.3 157 98.75 9 2.1 416 97.96 9 2.1 424 97.97 35 12.5 244 87.58 1 0.8 119 99.29 19 4.5 402 95.510 12 2.1 567 97.911 12 2.6 448 97.412 2 0.4 453 99.613 1 0.2 429 99.814 2 0.5 441 99.515 27 0.9 2920 99.116 16 0.5 3147 99.517 16 0.6 2480 99.418 2 0.1 2223 99.9
14
8
10
12
14
onerrors, %
2
4
6
Link
identificatio
0
0 2 4 6 8 10 12 14 16 18
Urban Urban and suburban
Urban, suburban and rural Suburban and rural
Fig. 3. Road link identification errors in different environments
TABLE III
SUMMARY OF LINK IDENTIFICATION ACCURACY ESTIMATION
Environment Correct link 95% confidenceidentification (%) interval
Mixed: urban, suburban, rural 97.9 0.9Urban 97.7 2.7Overall 97.8 0.9
the road link connectivity analysis is not used in the method,
and since no post-processing is involved, the method is
deemed useful in the situations when:
• digital road network map contains inaccuracies, such
as road curves interpolated as linear segments, missing
links, links outdated due to maintenance works or new
roads constructed;
• the matching needs to be done on-line and conserva-
tive computational overhead is important, e.g. in large-
scale solutions concurrently processing map-matching
requests from multiple traffic participants.
It should be noted that, since presumably inaccurate map is
used for mapping, the result of matching could hardly be
used for correcting positioning errors.
Based on the analysis of the results of map-matching, the
observed link identification errors can be categorized into the
following groups:
a) Missing a turn at a crossroad: Given equal distances
to the previous and the new road link, the method, due to
its Bayesian nature, will favor the previous road link. If
due to positioning errors, the estimated heading is closer
to the previous road link, the error of missing a turn is
likely. However, as soon as the distance to the previous link
increases, the correct new link is identified.
b) Errors at Y-junctions: At or near Y-junctions, the
heading of the road links differs insignificantly. If, due to
positioning errors, the estimated distance to an incorrect
link is smaller, the link identification error is likely. This
is sometimes observed near exit or entrance ramps until the
distance between road links becomes sufficient.
1496
c) Failing to recognize parking lots and correct lanes:
As mentioned above, the implementation ignores parking
areas and the directions of one-way lanes; as a result,
incorrect link identification is likely.
Many of these errors are caused either by the bias or
noise in the GPS positioning measurements. The results
of the experiments indicate that the method tolerates noise
in the positioning information well: the impact of random
deviations from the real position is smoothed during the
recursive Bayesian estimation process. However, the bias in
the measurements may reduce the link identification accuracy
dramatically: if the bias is of the same order as the distance
between closest links (with similar heading), the method will
eventually assign the measurements to incorrect link.
In further work, the method needs to be equipped with the
possibility to tolerate the bias in positioning measurements,
so that the above errors would be reduced. For instance,
Kalman filter can be used to estimate and correct the bias.
However, when positioning bias is too high (in some tests,
the bias exceeding 50 m was observed), the use of Kalman
filter will unlikely be able to compensate it.
Some of the errors may be avoided by delaying the map-
matching decision. For instance, if the identified road link
a (e.g. highway) changes for a second to b (e.g. exit ramp)
and then again returns to a, then the road link b is likely
to be identified incorrectly and can be filtered. However,
such delay would represent a form of post-processing, and
hence would limit the applicability of the method in online
applications.
In future work, some elements of topological analysis can
be used to assign the road link switching probabilities. In par-
ticular, near the end of a link higher switching probabilities
can be assigned to the adjacent links. If the information about
adjacent links is easily retrieved from the map database, the
computational overhead should be minimal. This should be
applied with care, however, since several road links may be
passed in the interval between consequent positioning mea-
surements (e.g. an approach to a roundabout, the roundabout,
and an exit from the roundabout), and analyzing adjacent
roads only may therefore be insufficient. Also, the benefit
of such connectivity analysis depends on the map accuracy,
and hence the influence of the connectivity on probability
assignment should be limited to tolerate inaccuracies in the
map.
Provisions should be made to accommodate the boundaries
of parking lots and other suitable for driving areas into the
analysis, and to equip the implementation with the possibility
to distinguish among opposite lanes in the maps.
VI. ACKNOWLEDGMENTS
The author would like to thank anonymous reviewers, as
well as colleagues from the Sensor Data Fusion and Appli-
cations project for their valuable comments and suggestions.
REFERENCES
[1] M. A. Quddus, W. Y. Ochieng, and R. B. Noland, “Currentmap-matching algorithms for transport applications: State-of-theart and future research directions,” Transportation Research
Part C: Emerging Technologies, vol. 15, no. 5, pp. 312–328,Oct. 2007. [Online]. Available: http://www.sciencedirect.com/science/article/B6VGJ-4P2JCX9-1/2/21d1ef73e42329087fc872e39ce78b1b
[2] W. Y. Ochieng, M. Quddus, and R. B. Noland, “Map-matching incomplex urban road networks,” Brazilian Journal of Cartography
(Revista Brasileira de Cartografia), vol. 55, no. 2, pp. 1–18, 2003.[3] M. Lehtinen, A. Happonen, and J. Ikonen, “Accuracy and time
to first fix using consumer-grade GPS receivers,” in 2008 16th
International Conference on Software, Telecommunications and
Computer Networks, Split, Dubrovnik, Croatia, 2008, pp. 334–340.[Online]. Available: http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=4669506
[4] A. Katasonov and M. Sakkinen, “Information quality assessment ofa yellow-pages location-based service,” in Computer Software and
Applications Conference, 2003. COMPSAC 2003. Proceedings. 27th
Annual International, 2003, pp. 320–326.[5] M. E. E. Najjar and P. Bonnifait, “Road selection using multicriteria
fusion for the road-matching problem,” IEEE Transactions on Intelli-
gent Transportation Systems, vol. 8, no. 2, p. 279, 2007.[6] J. Eriksson, L. Girod, B. Hull, R. Newton, S. Madden, and H. Balakr-
ishnan, “The pothole patrol: Using a mobile sensor network for roadsurface monitoring,” in Proceeding of the 6th international conference
on Mobile systems, applications, and services, 2008, pp. 29–39.[7] J. Kantola, M. Perttunen, T. Leppanen, J. Collin, and J. Riekki,
“Context awareness for gps-enabled phones,” in Proceeding of the
2010 Institute of Navigation International Technical Meeting (ION
ITM 2010), 2010.[8] D. Bernstein and A. Kornhauser, “An introduction to map matching
for personal navigation assistants,” New Jersey TIDE Center, 1996.[9] J. S. Greenfeld, “Matching GPS observations to locations on a digital
map,” in 81th Annual Meeting of the Transportation Research Board,2002.
[10] N. R. Velaga, M. A. Quddus, and A. L. Bristow, “Developing anenhanced weight-based topological map-matching algorithm for intel-ligent transport systems,” Transportation Research Part C, vol. 17,no. 6, pp. 672–683, 2009.
[11] K. Ghys, B. Kuijpers, B. Moelans, W. Othman, D. Vangoidsenhoven,and A. Vaisman, “Map matching and uncertainty: an algorithm andreal-world experiments,” in Proceedings of the 17th ACM SIGSPATIAL
International Conference on Advances in Geographic Information
Systems. Seattle, Washington: ACM, 2009, pp. 468–471.[12] J. S. Pyo, D. H. Shin, and T. K. Sung, “Development of a map
matching method using the multiple hypothesis technique,” in IEEE
Proceedings on Intelligent Transportation Systems, 2001, pp. 23–27.[13] L. Liao, D. J. Patterson, D. Fox, and H. Kautz, “Learning and inferring
transportation routines,” Artificial Intelligence, vol. 171, no. 5-6, pp.311–331, 2007.
[14] D. Yang, B. Cai, and Y. Yuan, “An improved map-matching algorithmused in vehicle navigation system,” 2003 IEEE Intelligent Transporta-
tion Systems, 2003. Proceedings, vol. 2, 2003.[15] M. A. Quddus, R. B. Noland, and W. Y. Ochieng, “A high accuracy
fuzzy logic based map matching algorithm for road transport,” Journal
of Intelligent Transportation Systems, vol. 10, no. 3, 2006.[16] W. Kim, G. I. Jee, and J. G. Lee, “Efficient use of digital road map
in various positioning for ITS,” in IEEE 2000 Position Location and
Navigation Symposium, 2000, pp. 170–176.[17] Y. J. Cui and S. S. Ge, “Autonomous vehicle positioning with GPS
in urban canyon environments,” IEEE transactions on robotics and
automation, vol. 19, no. 1, pp. 15–25, 2003.[18] D. Obradovic, H. Lenz, and M. Schupfner, “Fusion of map and
sensor data in a modern car navigation system,” The Journal of VLSI
Signal Processing, vol. 45, no. 1, pp. 111–122, Nov. 2006. [Online].Available: http://dx.doi.org/10.1007/s11265-006-9775-4
[19] C. E. White, D. Bernstein, and A. L. Kornhauser, “Some map matchingalgorithms for personal navigation assistants,” Transportation
Research Part C: Emerging Technologies, vol. 8, no. 1-6, pp. 91–108, 2000. [Online]. Available: http://www.sciencedirect.com/science/article/B6VGJ-417F93H-7/2/4e0354d6d2eb0925a0e55c13b7424162
1497