678 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO.

3, SEPTEMBER 2010
Intervehicle-Communication-Assisted Localization
Nabil Mohamed Drawil, Member, IEEE, and Otman Basir, Member, IEEE
AbstractVehicle localization is a key issue that has recently
attracted attention in a wide range of applications. Navigation,
vehicle tracking, emergency calling, and location-based services
are examples of emerging applications with a great demand for
location information. The Global Positioning System (GPS) has
been the de facto standard solution for the vehicle-localization
problem. Nevertheless, GPS-based localization is inaccurate and
unreliable due to GPS inherent poor performance in vertical
positioning and the prevalent horizontal movement, in addition to
anomalies caused by line-of-sight occlusions and multipath issues
in urban canyons. Although augmenting GPS localization with
inertial sensory data has demonstrated signicant performance
improvements, there remain situations that give rise to degraded
localization accuracya deciency that many applications cannot
tolerate. In this paper, we propose intervehicle-communication-
assisted localization, a localization technique that takes advantage
of the emerging vehicle ad hoc networks environments. Communi-
cation among vehicles is utilized to compute a relative vehicle lo-
cation, the integration of which with motion information and GPS
location estimates leading to highly accurate vehicle localization.
This proposed localization technique is tested in various simulated
road-segment scenarios. It is evident from the simulation results
that intervehicle communication has the potential to lead to the
improvement of the robustness and accuracy of vehicle-location
estimation.
Index TermsError detection, Global Positioning System
(GPS), localization errors, location estimate, optimization, vehicle
ad hoc network (VANET).
I. INTRODUCTION
V
EHICLE-location information is vital in many trans-
portation applications, including location-based services
(LBSs), navigation, emergency response, and assisted driving.
As a result, interest in vehicle localization has signicantly
grown in the last decade (see, e.g., [1][9]). Such emerging
systems cannot deliver on their requirements without accurate
localization. For example, in emergency calling (eCall) sys-
tems, when an accident occurs, the location information of the
vehicles involved in the accident can be used to notify the
nearest emergency center to deploy the required emergency
response. Furthermore, in LBSs, it is not possible to provide
the desired services without knowing the accurate location of
the vehicle that requests the service [10], [11].
GPS is a localization technology that is universally avail-
able. It has been the de facto standard solution for vehicle
Manuscript received March 31, 2009; revised December 16, 2009 and
April 6, 2010; accepted April 10, 2010. Date of publication May 17, 2010;
date of current version September 3, 2010. The Associate Editor for this paper
was S. C. Wong.
The authors are with the Department of Electrical and Computer Engi-
neering, University of Waterloo, Waterloo, ON N2L 3G1, Canada (e-mail:
[email protected]; [email protected]).
Color versions of one or more of the gures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TITS.2010.2048562
localization in a wide range of transportation applications.
However, GPS still suffers from accuracy and reliability issues,
to the extent that it fails to meet basic requirements, e.g., those
identied by the Federal Communications Commission (FCC)
[12]. For instance, a GPS receiver as a stand-alone localization
means can suffer from up to 50 m of positioning inaccuracya
particularly prevalent behavior in multipath environments
(urban canyon areas) [13]. Add to that issue the potential oc-
clusion of the GPS signal line of sight, e.g., in tunnel passages,
which often results in signal drop. These performance issues
have motivated many researchers to consider augmenting GPS
localization with other localization means, e.g., dead reckon-
ing and inertial navigation. Signicant accuracy improvements
have been reported for short GPS signal outage and intermittent
multipath conditions [14][18]; However, long GPS signal
outages and severe multipath situations still pose signicant
challenges for accurate and reliable vehicle localization.
In this paper, we propose to further the multimeasure-
ment approach to enhance localization in environments where
vehicle-to-vehicle communication is supported, i.e., vehicu-
lar ad hoc network (VANET) environments. We maintain
that, through intervehicle-communication-assisted localization
(IVCAL), knowledge sharing of potentially inaccurate vehicle
location, in conjunction with that of relative distances between
the vehicle and a selected set of vehicles in its vicinity, can be
used to achieve robust and accurate localization.
The remainder of this paper is organized as follows.
Section II presents some basic vehicle-localization techniques.
Section III compares different vehicle-localization methods
that use more than one source of location information.
Section IV introduces the proposed multimeasurement scheme
as a Kalman lter (KF) formulation. This formulation em-
ploys new optimization and updating schemes to estimate the
vehicles location. Section V presents experimental results to
demonstrate the performance of the proposed scheme in esti-
mating the vehicles location. Section VI provides concluding
remarks and recommendations for future work.
II. BACKGROUND
GPS is a positioning system that has been developed and
operated by the U.S. Department of Defense [19]. A GPS
system is formed from a network of satellites that transmit con-
tinuous coded information, which makes it possible to identify
locations on Earth by measuring distances from the satellites;
moreover, the receiver also has the ability to obtain information
about its velocity and direction. A GPS network consists of
24 satellites that are arranged in six orbital planes so that, at
any given time, a minimum of ve satellites can be observed
by GPS receivers at any location in the world. Different types
1524-9050/$26.00 2010 IEEE
DRAWIL AND BASIR: INTERVEHICLE-COMMUNICATION-ASSISTED LOCALIZATION 679
Fig. 1. Differential GPS receivers.
of GPS receivers have been developed for many applications
according to the accuracy required.
Basic GPS receivers often have multiple radio channels so
that the receiver can observe multiple GPS satellites at once
and obtain a pseudorange measurement from each satellite
signal. Leva [20] and Hoshen [21] show two different tech-
niques by which a GPS receiver can compute its location from
four pseudorange measurements and the minimum number of
measure localization in three dimensions. However, the pseudo-
range measurement is prone to many errors, and hence, it has a
typical accuracy range of 1050 m.
GPS errors can be categorized as either global or local errors.
Global errors are produced during satellite signals that originate
in and travel to the Earth, whereas local errors are receiver
surrounding environment dependent. The extent of the effect
of global errors on receiver measurements varies from one
geographical area to another due to ephemeris, ionospheric
delays, tropospheric delays, and satellite clock errors. Local
errors may arise as a result of multipath effects due to the lack of
line of sight or due to the receiver hardware itself. More details
about global and local GPS measurement errors are given in
[22]. The total GPS measurement error can be modeled as
follows:
GPS
err
= e
eph
+ e
iono
+ e
trop
cT + e
mp
+ e
rcvr
(1)
where e
eph
is the satellite ephemeris error, e
iono
is the
ionosphere error, e
trop
is the troposphere error, c is the speed
of light, T is the receiver clock error, e
mp
is the multipath error,
and e
rcvr
is the error caused by the receiver hardware. In dense
urban areas, the E
tot
is dominated by the e
mp
[23].
It is possible to minimize global errors and obtain more accu-
rate location measurements by using a differential GPS (DGPS)
receiver. A DGPS consists of two receivers that observe the
same GPS satellites (see Fig. 1). One of these receivers is
stationary, and the other one, which is used to determine the
vehicle location, roves. The stationary receiver, which resides
at a known location, obtains the pseudorange from the satellite
signals. These pseudorange measurements are used to compute
a global error by comparing the GPS location of the stationary
receiver with the receivers known location. The stationary
receiver computes and transmits a global error correction to the
roving receiver so that it can correct its location estimate. In
the best cases, the DGPS-based accuracy improvement can be
as good as tens of centimeters. However, this type of solution
has drawbacks, e.g., the cost of communication between the
stationary receiver and the roving receiver(s) and the cost of
the hardware. Moreover, DGPS receivers must be under the
coverage of the same GPS satellites to be correlated, i.e., having
the same global error. Hence, this requirement constrains the
movement of the roving receiver to be within a bounded area.
Furthermore, DGPS cannot correct for local errors, because the
multipath effects occur in the vicinity of the roving receiver.
Another method for achieving improved localization is the
dead-reckoning system, which has been adopted in some appli-
cations. In this technique, the location estimation is computed
based on how far the vehicle has moved from a known place,
given the directions and distances traveled over small periods
of time. This technique is simple and inexpensive; therefore,
it has been the choice for many applications. One crucial
disadvantage of this method is that errors in the direction and/or
distance measurements affect the nal location estimation. Fur-
thermore, because the measurement errors accumulate over the
total period of time, such a method is recommended for use over
short periods of time.
III. RELATED WORK
GPS and the dead-reckoning system are the commonly used
techniques in vehicle-navigation systems. Each technique has
a complementary capability to the other one; therefore, in-
tegrating the two methods has emerged as one approach for
achieving superior performance. A large body of research has
been motivated by this idea from different points of view.
For instance, a GPS receiver and a dead-reckoning system
can be combined in one navigation system so that the GPS
corrects the accumulated error caused by the dead-reckoning
system; when the GPS measurement is not available, the dead-
reckoning system estimates the location by using sensors, e.g.,
wheel odometers, a ux-gate compass, a gyroscope, and an
accelerometer [14]. Similarly, it is possible to integrate the GPS
reading with the inertial navigation system (INS) by means of
a KF [16]. The computational complexity of the extended KF
(EKF) can be reduced by preprocessing the INS measurements
and inputting them to the KF as a linear component. However,
preprocessing the INS measurement adds to the computational
cost of the solution. On the other hand, a real-time kinematic
(RTK) GPS is proposed using an EKF, [15], where the GPS
latency is identied as the time required for the satellite signals
to travel to Earth and the time required for the computation of
the location or one or more of the GPS satellites to change
while the signals are received. These delays may corrupt the
synchronization of the readings of the sensors, particularly
when they are used in the control loop of a moving vehicle to
estimate its location using a high-precision system.
The nonlinearity of the models used in GPS/INS fusion
systems constitutes prevalent difculty in the balance between
680 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010
improving the localization accuracy and decreasing the com-
putation cost. Approximation of nonlinear lters has been
approached, e.g., in [24] and [25], and achieved a remarkable
improvement in the accuracy and cost of the localization
process. However, to the best of our knowledge, the localization
performance in severe multipath environments has not been ad-
dressed in the reported research. Such environments drastically
affect on the GPS measurements and contribute the most to the
overall localization error.
An articial neural network (ANN) is chosen as a tool for
recognizing patterns of errors and noises in the DGPS/INS
measurements [17]. The work in [17] is quite similar to that
in [15] in that preprocessing operations are performed on the
measurements before they are fused.
Integrating the INS of a dynamic model with a DGPS is
also investigated in [18]. To deal with the nonlinearity of the
dynamic model, an EKF is used. Localization accuracy of 0.9 m
on 100-m driving track was reported for situations where the
system relies on the dynamic model more than it does on the
GPS measurements. The multipath effect is not addressed as
the experiment was conducted in an open-space environment.
It can thus be concluded that, even with the most advanced
and expensive GPS/DGPS, it is essential to integrate GPS
measurements with other measurements, e.g., those from INS,
to improve the accuracy of the localization process. None of the
aforementioned solutions can cope with the problem of satellite
signals and/or the line of sight loss; therefore, it is imperative to
identify other sources of information that can help with the goal
of improving location estimates in the presence of GPS signal
blockage or with large measurement errors.
Map-matching methods have been the choice of many re-
ported research work as a localization method augmented with
GPS/INS localization systems, e.g., in [26]. Map matching
utilizes the available digital maps of roads stored in a geo-
graphic information system (GIS). The roads in a GIS are
represented in the form of segments that comprise a network
of polylines. The outputs of the map-matching process are
the segment identication number of the closest road to the
location estimate obtained by the GPS/INS system and the
relative position of the vehicle along this segment. The usage
of map matching, nevertheless, in a compact network of roads
might cause persisting localization errors, particularly in urban
environments where the GPS/INS estimate is not reliable, not
to mention the possible shift between the map and the ground
truth.
Some research has focused on GPS-less techniques for local-
ization [27], [28]. These techniques mainly rely on estimating
the distance between two objects (i.e., a sender and a receiver
in a wireless network), which is a method inspired by the
localization approaches used in cellular networks [29][31].
These approaches are based on radio-location methods, e.g.,
time of arrival (TOA), angle of arrival (AOA), received signal
strength (RSS), or time difference of arrival (TDOA) [32]. Once
the distances between an object and other reference points are
obtained, trilateration can be performed to estimate that object
location. However, radio location methods still suffer from two
types of errors: 1) the multipath effect or non-line-of-sight
(NLOS) errors and 2) hardware-measurement errors.
In [28], the NLOS is mitigated using the algorithm proposed
in [31]. Although it is reported that the system attains enough
stability and location accuracy, despite the error in the range
distance obtained using the TOA method, the system does not
suit vehicle applications (particularly in high-speed situations)
and cannot scale up due to its inherent intercommunication
overhead. On the other hand, this method is appropriate in slow
mobility applications.
Under the assumption of a VANET environment, a localiza-
tion method is introduced in [33] to localize vehicles that are
unequipped with GPS receivers or those whose location cannot
be determined, because the satellite signals are lost, e.g., in
a tunnel. In this method, vehicles that are not equipped with
GPS determine their own locations by relying on information
that they receive from vehicles that are equipped with GPS.
Depending on the radio range, every two vehicles can commu-
nicate within a limited distance; therefore, they can measure the
distances between themselves using one of the radio-location
methods presented in [32]. By nding three neighbors that
are the closest to the unequipped vehicle, its position can be
computed using trilateration. Other cases, e.g., when fewer than
three neighbors are available, are thoroughly discussed in [33].
It is reported that, when as many as 40% of the total number
of vehicles in the network are GPS equipped, the system can
be optimal; nevertheless, no mention is made of the level of
accuracy or even the type of GPS receiver used on the equipped
vehicles.
Developing a collaborative and intelligent sensory localiza-
tion system for vehicles is the aim of this paper. Such a system
takes advantage of some of the reported research to improve the
robustness and accuracy of location estimate under local and
global measurement errors.
We maintain that the use of a KF to integrate GPS reading
with measurements fromother sensors constitutes a viable strat-
egy. However, it is well understood that such a KF formulation
will continue to be challenged with issues, e.g., satellite signal
losses and multipath effects. As such, we propose to com-
plement this strategy, assuming a VANET environment, with
location estimations that are computable based on intervehicle
communication among the VANET vehicles.
IV. INTERVEHICLE-COMMUNICATION-ASSISTED
LOCALIZATION
The approach proposed in this paper introduces a novel tech-
nique where the KF and the intervehicle communication system
collaborate to increase the localization robustness and accuracy
for every vehicle in the network. The two main components that
interconnect the intervehicle communication systemand the KF
are given as follows: 1) the multipath detection unit (MDU),
which has the ability to detect the existence of multipath effects
in the output of the KF, and 2) the localization-enhancement
unit (LEU), which obtains the neighbors information from
the intervehicle communication system and feeds back an op-
timized location estimate to the KF. The diagram of the new
approach is depicted in Fig. 2.
It is assumed that every vehicle of the VANET is equipped
with a GPS receiver, an INS, and a VANET transceiver.
DRAWIL AND BASIR: INTERVEHICLE-COMMUNICATION-ASSISTED LOCALIZATION 681
Fig. 2. Block diagram of the proposed technique.
Fig. 3. VANET.
Each vehicle uses a GPSINS fusion unit (GPS/INS-FU) to
integrate measurements from the INS and the GPS using a KF.
As depicted in Fig. 3, the intervehicle communication system
of each vehicle extracts information pertinent to the location
estimates of vehicles in its vicinity. This information includes
the distance between the vehicle and each of its neighbors, the
location estimates of the neighbors, and the level of uncertainty
in the location estimates of the neighbors. This information
compensates for the loss of satellite signals and helps correct
for errors that are caused by the multipath effect (assuming that
not all vehicles of the network suffer from a multipath effect at
the same time). Vehicles with the smallest uncertainty in their
location estimate are considered to be location anchors.
Prior to providing the details of the MDU and LEU, we rst
introduce a model for fusing the GPS and the INS measure-
ments using a KF.
A. GPS/INS-FU
Similar to [14][18], a KF is employed to fuse GPS and
INS data to improve localization accuracy. Without loss of
generality, xed speed and direction are assumed in the for-
mulation. As such, the location of a VANET vehicle v
i
V =
{all vehicles in VANET} can be modeled as
x
i
k+1
= x
i
k
+t x
i
k
(2)
where k is a time slice index, the state variable x
i
k
represents
the exact vehicle location at time t
k
, and t is the sampling
period of time.
x
i
k
R
2
represents the vehicle velocity, which is obtained
from the INS as follows:
x
i
k
=
_
x
y
_
i
k
= S
i
k

_
cos
i
k
sin
i
k
_
(3)
where S
i
k
R represents the vehicle speed, and
i
k
[0

, 360

]
represents the vehicle direction with respect to the global axis.
In addition, each vehicle has a location estimate obtained by
its GPS receiver, i.e.,
z
i
GPS,k
=
_
x
i
y
i
_
GPS,k
(4)
where x
i
and y
i
R represent v
i
s global coordinates. A
discrete KF is used to integrate these two measurements as
follows:
x
i
k+1
=Ax
i
k
+B x
i
k
+w
i
k
z
i
k+1
=Hx
i
k+1
+

i
k+1
(5)
where
A =
_
1 0
0 1
_
, B =
_
t 0
0 t
_
, H =
_
1 0
0 1
_
where x
i
k+1
is a 2 1 vector that represents the state or location
of the vehicle v
i
at time t
k+1
, given the location and the velocity
at time t
k
, and z
i
k+1
is a 2 1 vector that represents the
expected GPS reading at t
k+1
. The matrices A, B, and H are
2 2 transition matrices that relate the current state and current
input to the next state and the expected state to the next GPS
measurement, respectively. The process and the measurement
noise are represented by 2 1 random vectors w
i
k
and

i
k
,
respectively. They are assumed to be independent and white
with normal probability distributions. We have
w
i
N(0, Q)

i
N(0, R).
682 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010
Q and R are 2 2 matrices that represent the process noise
covariance and the measurement noise covariance, respectively.
These two matrices resemble the covariance of the INS error
and the error of the GPS receiver when no multipath effect is
present.
It is practically impossible to separate these two types of
noise from the measurements; therefore, the system in (5) is
implemented as follows:
x
i
k+1/k
=Ax
i
k/k
+B x
i
k
z
i
k+1/k
=Hx
i
k+1/k
(6)
where x
i
k+1/k
is dened as an a priori state estimate, and
x
i
k/k
is dened as an a posteriori state estimate, which can
be calculated given the GPS measurement at t
k
, as shown in
the following discussion. Accordingly, two types of estimation
errors are introduced: 1) an a priori estimate error 2 1 e
i
k+1/k
and 2) an a posteriori estimate error 2 1 e
i
k+1/k+1
. We have
e
i
k+1/k
=x
i
k+1
x
i
k+1/k
e
i
k+1/k+1
=x
i
k+1
x
i
k+1/k+1
. (7)
The a priori estimate error covariance is then
P
i
k+1/k
= E
_
e
i
k+1/k
e
i
k+1/k
T
_
(8)
and the a posteriori error estimate is
P
i
k+1/k+1
= E
_
e
i
k+1/k+1
e
i
k+1/k+1
T
_
. (9)
The goal of the KF is to compute an a posteriori estimated
state x
i
k+1/k+1
in two cases, as in (10). If no multipath effect is
detected, the KF uses an a priori estimated state x
i
k+1/k
and a
weighted difference between an actual measurement z
i
k+1
and
an estimated measurement z
i
k+1/k
as in (11). However, if a
multipath effect is detected, the KF uses the output of the LEU
( x
i
k+1/k+1
), which will be discussed later. We have
x
i
k+1/k+1
=
_
x
i
k+1/k+1
, no detection of multipath
x
i
k+1/k+1
, detection of multipath
(10)
x
i
k+1/k+1
=x
i
k+1/k
+G
i
k+1
_
z
i
k+1
z
i
k+1/k
_
. (11)
A difference of zero between the predicted measurement
and the actual measurement in (11) means that the measure-
ments are in complete agreement. This difference is called the
measurement residual. The residual reects the discrepancy
between the estimated and the actual measurement.
The 2 2 matrix G in (11) is the estimate error covariance
in (9). This minimization can be accomplished if (11) is substi-
tuted in (7), the expectation of (9) is performed, the derivative
of the trace of the result with respect to G is taken, and G
is solved for after the result is set to be equal to zero. G has
different forms, one of which is given as follows:
G
i
k+1
= P
i
k+1/k
H
T
_
HP
i
k+1/k
H
T
+R
_
1
. (12)
This blending factor is the Kalman gain. In addition, the
a priori and a posteriori error covariances can be dened as
P
i
k+1/k
=AP
i
k/k
A
T
+Q (13)
P
i
k+1/k+1
=
_
I G
i
k+1
H
_
P
i
k+1/k
. (14)
With the aforementioned equations, all the quantities that are
required for estimating the system state at t
k+1
are determined.
Equations (6), (10), and (12)(14) comprise the KF recursive
equations.
B. MDU
The MDU enables vehicles to detect the presence of a multi-
path effect in their location estimates ( x
i
k+1/k+1
) to trigger the
LEU to minimize the localization error that is mostly affected
by the e
mp
depicted in (1).
The multipath effect adds noise to the location estimate and
contaminates the randomness of the measurement error so that
it cannot be represented by the matrix R of the KF. Accord-
ingly, the KF is not optimumin cycles when the multipath effect
is present.
Equation (15), which is derived based on (11) and is shown
below, denes the discrepancy ( ) between the estimated loca-
tion in the time update stage, i.e., (6), of the KF and the actual
measurement, i.e.,

i
k+1
= z
i
k+1
Hx
i
k+1/k
. (15)
This discrepancy measure reects the divergence between
the GPS receivers measurement and the location estimate. It
can be used to determine whether a multipath effect is present.
Traditionally, the discrepancy value is normalized using the
observation covariance matrix of the KF and then tested against
a predetermined threshold to detect multipath signals in the
GPS measurement [34]. However, because a vehicle is expected
to be driven in a wide range of environmental conditions, it
is impractical to assume that the discrepancy patterns to be
linearly separable. Therefore, a neural network is proposed
to classify these patterns. Training the neural network on a
wide range of divergence patterns, including those obtained
for various multipath contaminated measurements, produces a
classier that can identify multipath affected measurements.
A feedforward backpropagation network (FFBN) is chosen in
this paper. FFBN falls under the supervised learning category
of ANNs. As argued in [35], a feedforward neural network
with a sigmoid transfer function in the hidden layer and a
linear transfer function in the output layer can approximate
any nonlinear function to a high degree of accuracy, provided
that the hidden layer employs a sufcient number of neurons.
Furthermore, feedforward neural networks lend themselves to
embedded hardware implementations [36].
In this paper, the input to the FFBN is the magnitude of the
discrepancy vector
i
k+1
, because the aim is to detect the
presence of a multipath effect in any of the vector components.
The FFBN is constructed to form three layers (input, hidden,
and output), as shown in Fig. 4. One neuron is used in the
input layer to pass the input pattern to the neurons of the
DRAWIL AND BASIR: INTERVEHICLE-COMMUNICATION-ASSISTED LOCALIZATION 683
Fig. 4. Neural network for detecting the multipath effect.
hidden layer. Hyperbolic tangent sigmoid transfer functions are
used in 50 neurons of the hidden layer. One neuron consti-
tutes the output layer, which provides the MDU decision after
thresholding a weighted linear combination of the output of the
hidden neurons. The number of neurons in the hidden layer was
empirically chosen; therefore, the FFBN can accurately detect
multipath signals.
The FFBN was trained using discrepancy samples randomly
generated in two different simulated environments: 1) open
area and 2) severe multipath regions. Each region was simu-
lated using specic additive white Gaussian noise in the GPS
readings. This approach is tractable in practice by gathering
discrepancy samples using a GPS/INS technique in an open-
sky environment and a city-center region saturated by high
buildings.
The performance of the trained FFBN was tested by exam-
ining various discrepancy values. The discrepancy values used
in the testing were obtained from four simulated environments
with different standard deviations of GPS noise: 1) 15 m;
2) 60 m; 3) 90 m; and 4) 150 m. Fig. 5 shows the performance
of the neural network. It is shown that the MDU can detect mul-
tipath signals, even if they are not severe (i.e., white Gaussian
noise with 60 m of standard deviation).
C. LEU
To cope with multipath signal deterioration, each vehicle
communicates with its VANET neighbors, which are not af-
Fig. 5. Performance of the MDU in detecting multipath signals. (a) Strength
of the multipath effect versus time. (b) Resulting discrepancy values. (c) ANN
response. (d) Final output of the MDU.
fected or are less affected by the multipath disturbance. Thus,
it is important for the system to nd a measurement that can
indicate the accuracy of the neighbors. This measurement will
help a vehicle select the best three neighbors, which it will
684 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010
consider to be location anchors
1
that the vehicle can use to more
accurately localize itself.
In this paper, the accuracy of a vehicles location is repre-
sented by the variation in the discrepancy between the time
update estimate of the KF and the GPS receiver measure-
ment. Each vehicle maintains a window that comprises recent
discrepancy values. The size of this window is such that the
uncertainty of the recent location estimate can be represented
by determining the variance in the values.
When a vehicle does not experience any multipath effect, the
accuracy of its GPS will be high, and the discrepancy values
will be low and comparable in value. As a result, the variation
in the discrepancy values will also be small. In other words, this
vehicle will have a small uncertainty about its location, and it
can therefore be used as a location anchor for its neighbors, and
vice versa.
Oftentimes, after experiencing a multipath effect, the KF es-
timate needs cycles of updates before it converges to a low-error
vehicle-location estimate. Every vehicle maintains a record of
the previous discrepancy values; therefore, the uncertainty of
the location estimate is expected to remain large for an extended
period of time, even after a vehicle leaves a multipath region.
Vehicles that experience this condition are not used as location
anchors.
A vehicle v
i
V with uncertain location estimate uses
the information that it obtains from three location anchors
(v
a
, v
b
, v
c
= {vehicles in the coverage area of v
i
}) to cor-
rect its location estimate using the following least squares
formulation.
Let x
i
k+1/k+1
be the KF location estimate of vehicle v
i
computed at time t
k+1
. x
a
k+1/k+1
, x
b
k+1/k+1
, and x
c
k+1/k+1
are the locations estimates of the three location anchors in .
At time t
k+1
, the MDU detects a multipath effect, and hence,
v
i
s location is optimized as follows:
x
i
k+1/k+1
=arg min
x
i
k+1/k+1

jI
_
d
i,j

_
_
_ x
i
k+1/k+1
x
j
k+1/k+1
_
_
_
_
2
(16)
where I = {a, b, c} is the index set of the selected location
anchors, x
j
k+1/k+1
are the locations of the selected anchors, d
i,j
is the distance between the vehicle V
i
and the anchor V
j
[32],
x
i
k+1/k+1
is the optimized location estimate, and x
i
k+1/k+1
is
the initial location of the vehicle V
i
.
The aforementioned optimization problem is nonlinear, con-
tinuous, and convex. It can therefore be solved with the
NelderMead method [37]. This method is fast, because it can
be implemented without calculating the gradient vector or the
Hessian.
However, the ideal case of the optimization problem, in
which the error in the location and distance estimates of the
location anchors is zero, produces one global minimum. Fig. 6
shows a contour depiction of the values of the objective function
using different possible values for the solution of the optimiza-
1
In this paper, the localization is considered to be in two dimensions; if three
dimensions are required for localization, then the minimum number of anchors
would be four.
Fig. 6. Contour representation of the optimization problem in the ideal case.
Fig. 7. Example of the solution for the optimization problem in the ideal case.
tion problem, which is the location of the location-uncertain
vehicle (LUV). From a geometric point of view, the solution of
such an optimization problem is the intersection point of three
circles. These circles represent the distances between the LUV
and the reference points; the circles are centered at the locations
of the reference points, and the radius of every circle is equal to
the distance between the LUV and the center of that circle, as
shown in Fig. 7. Obviously, as depicted in Fig. 6, there is only
one global minimum for the optimization problem, which is the
actual location of the LUV.
Typically, nding the ideal case in an experiment is hard, let
alone in the real world, because every vehicle has an error in
its location estimate, not to mention the error in the distance
estimates (d
i,j
). Such localization errors result in two, three, or
more local minima in the optimization problem, representing
signicant drift from the actual location of vehicle.
DRAWIL AND BASIR: INTERVEHICLE-COMMUNICATION-ASSISTED LOCALIZATION 685
Fig. 8. Small errors in the location estimates of the reference points create
more than one solution for the optimization problem around the actual location
of the LUV.
Fig. 9. Two local minima as a solution for the optimization problem in the
real case.
For example, Fig. 8 shows an LUV that attempts to improve
its location estimate by communicating with three other vehi-
cles that are more accurate. These reference vehicles have error
values in their location estimates that cause the solution of the
optimization problem to fall a signicant distance from the ac-
tual location of the LUV. Fig. 9 shows the two solutions for the
optimization problem of this example. If such an optimization
problem has more than one solution, the actual location of the
LUV is often midway between these solutions [38], as depicted
in Fig. 8. Therefore, if these local minima can be found, their
centroid will be close to the actual location of the LUV.
Thus, we have to explore the area around x
i
k+1/k+1
, which
is the closest coordinate to the actual location that we know, to
nd more local minima, if there are any. This search technique
can be implemented through the division of the space around
x
i
k+1/k+1
into four parts and solving the optimization problem
four times with different initial conditions from each part. This
way, the rst initial condition for the optimization problem
is x
i
k+1/k+1
. Then, x
i
k+1/k+1
is used as a reection point to
Fig. 10. LEU searches for the local minima and computes the centroid point.
nd new other initial conditions. If all the initial conditions
lead to the same minimum point, then there is only one global
minimum solution, which is close to the actual location of the
LUV. Otherwise, the result will be a number of local minima
scattered around the actual location of the LUV, which can then
be found by computing the centroid of these minima.
Fig. 10 shows how the LEU behaves when it searches for
the local minima for the previous example shown in Fig. 9.
2
The rst initial condition ( x) is equal to x
i
k+1/k+1
, in which
the optimization technique nds the local minimum at the top
of the gure ( x
1
). The second initial condition ( x
2
) leads the
optimization technique to the previous local minimum ( x
1
).
However, the third and fourth initial conditions ( x
3
, x
4
) lead
the optimization technique to the local minimum at the bottom
of the gure ( x
2
).
The following equation depicts how the initial conditions can
be determined using the rst local minimum x
1
and the rst
initial condition x:
x
n+1
= x +J(n)( x
1
x)
where n = 1, 2, 3, x
n
s are the initial conditions, and J(n) is an
n 90

rotation matrix.
The last step is to determine the location estimate by the
computation of the centroid of the local minima x, which is
equal to x
i
k+1/k+1
in (10).
We can now summarize the steps performed in the LEU if a
multipath effect is detected.
2
In this part of the paper, the superscript and the subscript of x will not
represent the time and the vehicle index, because the concern here is about one
vehicle in one cycle of the KF. Instead, the subscript of the initial conditions x
represents the part from which the initial condition is selected, and the subscript
of the local minima xs represents an index for these minima.
686 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010
D. LEU Algorithm
When a vehicle detects that its location estimate has been
contaminated by the multipath effect, the following algorithm
must be performed to minimize the impact of the multipath
effect.
1) The LUV (v
i
) sends a message to its neighbors, announc-
ing that its location estimate needs to be corrected.
2) The LUVs neighbors (v
j
) reply by sending their
location estimates and their uncertainty evaluations.
3) The distances (d
i,j
s) between the LUV v
i
and every
neighbor (i.e., v
j
) are determined using one of the tech-
niques mentioned in [32].
4) According to the neighbors uncertainty evaluation, the
LUV chooses as reference nodes three neighbors with
much better accuracy
3
than the LUV.
5) If there are enough more accurate vehicles (three location
anchors), the LUV corrects its location estimate by using
the aforementioned optimization technique.
6) If there are not enough accurate vehicles (fewer than
three location anchors), then the inaccurate KF estimate
x
i
k+1/k+1
is ignored, and x
i
k+1/k+1
will be set equal to
the time update KF estimate x
i
k+1/k
.
7) The uncertainty of the LUV is updated. If there are
enough reference nodes, the worst uncertainty measure-
ment among them is used; if not, the inaccurate uncer-
tainty measurement is omitted.
Step 6 takes advantage of the small error caused by the INS
system over a short time and cancels the inaccurate location
estimate. The likelihood of not having enough anchors is re-
mote, because not all vehicles will simultaneously suffer from
the multipath effect. In addition, the surrounding environment
is different from one vehicle to another and quickly changes
with the vehicles movement. However, even if this situation
occurs, it will not last for long, which guarantees that the
location estimate will not be negatively affected using INS
measurements.
V. EXPERIMENTAL RESULTS
This section presents the implementation of the algorithms
and theoretical work explored in Section IV. First, the imple-
mentation of the KF without the IVCAL technique for a single
vehicle is described. Then, the effect of multipath signals on
GPS readings and the performance of the KF are demonstrated.
Next, the plan of the simulation scenarios and the experimental
steps for a group of vehicles in VANET are explained. This sec-
tion concludes with a discussion of the results and a comparison
of the different cases and scenarios.
A. KF Implementation Without IVCAL for One Vehicle
As discussed in this paper, inaccuracies of location estimates
can be attributed to INS measurement errors and GPS receiver
errors, as explained in Section II. It is evident in (3) that an
error in the INS measurement (vehicle speed or direction) will
3
A fraction of the LUVs uncertainty is set as a threshold, and the uncertainty
of every neighbor is compared with this value.
Fig. 11. Example of the localization technique using the KF to integrate
the INS measurements and the GPS receiver readings in an open area when

R
= 10 m.
affect the vehicle location estimate (i.e., the system state). The
KF formulations capture this error in the following process
covariance matrix:
Q =
2
Q
I.
A value of 0.5 m has been suggested in [39] and [40]
as an experimentally measured value for
Q
. GPS receiver
errors are represented by the measurement noise covariance
matrix R, i.e.,
R =
2
R
I
where
R
is the error standard deviation of the GPS location es-
timate (typically between 10 and 15 m in open-sky conditions).
However, R does not capture the GPS receiver measurement
error when the KF estimation is deteriorated because of a
multipath effect. In such cases, the error standard deviation of
the GPS receiver reading becomes between 100 and 150 m,
and this random change causes a signicant increase in the
uncertainty of the KF estimate. In the simulation, the sampling
rate of the GPS receiver reading is 1 Hz, which is the typical
frequency of a GPS receiver. Similarly, the INS measurements
are sampled at the same rate.
Fig. 11 shows an example of how the KF estimates the
locations of a vehicle that travels a distance of 5 km at a
speed of 50 km/h. Fig. 12 depicts the multipath effect on the
vehicle when it passes through a 500-m region with high-
rise buildings. This region is located between the 2000 m and
2500 m points of the total distance traveled. Fig. 13 shows a
magnied multipath effect in Fig. 12. It is shown in this gure
that the multipath effect is still present, even after the vehicle
has passed the multipath region. This result can be explained
by the KF memory. It will take the KF a few cycles to lter
out these error residuals. It is worth mentioning that, in the
DRAWIL AND BASIR: INTERVEHICLE-COMMUNICATION-ASSISTED LOCALIZATION 687
Fig. 12. Example of the localization technique using the KF to integrate the
INS measurements and the GPS receiver reading. The vehicle in this example
experiences a multipath effect for 500 m, in which
R
= 150 m.
Fig. 13. Example that shows that the KF is still inuenced by the multipath
effect, even after the vehicle has passed the multipath region.
case of intermittent multipath situations, the MDU can still
improve the localization performance of the GPS/INS fusion
unit by assigning less weight to the GPS measurement in such
situations.
B. Simulation Scenario
The simulation for this paper was implemented using
MATLAB7. The scenario covers a 5000-m (5-km) portion of a
straight road. Vehicles that travel this road experience different
local environments, e.g., an open area with no multipath effect
and an inner city area where high buildings cause a severe
multipath effect. The road consists of two lanes in one direction
for vehicles with different speeds. The right lane contains
vehicles that travel at 50 km/h, and vehicles in the left lane
travel at a higher speed of 60 km/h. The width of each lane
is 3 m, and vehicles are assumed to be driven in the center
of the lanes. The simulation uses 100 vehicles to comprise the
VANET, which are uniformly distributed on the 5-km portion
of the road. The simulation period is equal to the time required
for a vehicle to pass through 5 km at a speed of 50 km/h: 360 s.
As shown in Fig. 14, on the rst 2000 m of the road, the
vehicles travel through an open-area environment. They then
experience a multipath effect over a distance of 300 m. Another
region of open area follows for 400 m. The second multipath
area occurs for the next 400 m. Then, the vehicles travel through
an open area to the end of the road.
To maintain the connectivity of the VANET, the number of
vehicles on 5 km is kept constant at 100. When a vehicle passes
the 5-km point, a new vehicle enters that portion of the road
from the opposite side.
The purpose of setting up the scenario this way is to show
the multipath effect and how long it will last after a vehicle has
passed through the multipath area.
Next, the IVCAL technique is rst tested in the presence
of noise only in the GPS signal where the estimations of the
intervehicle distances are assumed to be noise free. Noisy esti-
mations of the intervehicle distances are then applied to IVCAL
to show how this type of noise affects on the performance of the
IVCAL technique.
C. Results
In the rst experiment, the VANET simulation was run using
a KF without the proposed technique, and then, the simulation
was again run with the IVCAL technique, in which zero noise is
assumed in the relative distances estimation. The coordinates of
a vehicle were captured as it traveled over 5 km. Fig. 15 shows
how the estimate of the vehicles location is drastically affected
and becomes unreliable when the vehicle passes through the
high-building regions. In Fig. 15, the dotted line represents the
real path of the vehicle. The dashed line represents the location
estimate using the KF. The solid line represents the location
estimate using the proposed technique.
1) Comparison of the Localization Error Statistics: The
statistics related to the localization error are provided in Table I.
These statistics were calculated by computing the localization
error LE
i
t
of v
i
in every second t, because the localization
technique estimates locations every second. Then, the mean and
the standard deviation of the localization errors were computed
for each different environment as follows:
LE
i
t
=
_
_
_x
i
t
k
x
i
k/k
_
_
_
2
Mean Error (i) =
1
N
t
ex

t=t
en
LE
i
t
Standard Deviation of the Error (i)
=
_
1
N
t
ex

t=t
en
_
LE
i
t
Mean Error (i)
_
2
_
1/2
688 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010
Fig. 14. VANET simulation scenario.
Fig. 15. Location estimate for one vehicle in the VANET scenario. The
vehicles real path is depicted by the dotted line. The localization implemented
using only the KF technique is depicted by the dashed line. The localization
implemented using the IVCAL technique is depicted by the solid line.
TABLE I
STATISTICS FOR LOCALIZATION USING THE KF
AND THE IVCAL TECHNIQUES
where t
en
, t
ex
represent the times of entering and leaving a
region, and N represents the number of samples in a region.
In Table I, the rst column shows the type of environment
that the vehicle experiences, the second column lists the length
of the distance traversed in that environment, and the third and
fourth columns give the mean error and the standard deviation,
respectively, of the error-in-location estimate for the same
vehicle in the two experiments.
It is shown that the localization technique using the KF to
fuse only the GPS and INS measurements can be reliable in
the rst open-area region. However, the increase in the standard
deviation and the mean of the localization error in the rst
high-building region proves the unreliability of the location-
estimate technique by using only the fusion of the GPS and
INS measurements in the presence of the multipath effect.
Moreover, the multipath effect still appears in the mean and the
standard deviation of the localization error during the second
open-area region, even after the vehicle has left the rst high-
building region. In the second high-building region, the mean of
the localization error dramatically increases, because the local-
ization error has been large since the beginning of this region.
The multipath effect that remains in the location estimate from
the previous open area causes an error in the location estimate
at the beginning of this region. Again, the standard deviation
of the localization error during the second high-building region
shows the instability caused by the multipath effect. The KF
nally brings the localization error back to almost the same
accuracy as at the beginning of the rst open-area region.
However, it takes very long to minimize the error in the last
open area: The vehicle has traveled almost 1.3 km before the
multipath effect is eliminated.
Although the performance of the two techniques is similar
in the rst open-area region, the localization mean error and
standard deviation error are dramatically reduced with the
IVCAL technique during the remainder of the scenario.
The cumulative distribution functions (cdfs) of the localiza-
tion error relevant to IVCAL and GPS/INS fusion techniques
are plotted and compared in Fig. 16. As expected, the more
the location information available, the greater the localization
accuracy. As shown, the GPS/INS fusion produced 66% of the
DRAWIL AND BASIR: INTERVEHICLE-COMMUNICATION-ASSISTED LOCALIZATION 689
Fig. 16. CDF of the location estimate error for the IVCAL and GPS/INS-FU
techniques.
Fig. 17. Performance measures. The average ratios of the total error in
the location estimate using the KF (GPS/INS) technique to that using the
IVCAL technique are superimposed on the histogram of the time-performance
measure.
vehicle location estimates with errors less than 5 m, where 82%
of the vehicle location estimates had errors less than 5 m using
the IVCAL technique.
As shown in Fig. 17, the simulation results when vehicles use
the IVCAL technique can be summarized as follows.
The IVCAL technique outperformed the GPS/INS fusion
in 14 out of 100 vehicles during more than 90% of the
Fig. 18. Effect of the noisy intervehicle distances estimation on the localiza-
tion error of IVCAL.
simulation time and minimized the location estimate error
by 53% on the average.
The localization error, on average, was minimized by
37.5% in 65 out of 100 vehicles when the IVCAL tech-
nique outperformed the GPS/INS fusion in 50%90% of
the simulation time.
Although the IVCAL technique outperformed the
GPS/INS technique in only 20%40% of the simulation
time in seven out of 100 vehicles, about the same
amount of localization error was produced by the two
techniques.
In other words, the IVCAL technique signicantly improved
the vehicle-localization process. In the worst case, IVCAL pro-
duced the same localization error produced by the conventional
technique.
2) Effect of the Relative Distance Uncertainty: The TOA
technique [32] is assumed to be used in the simulations to
estimate relative distances among vehicles in IVCAL. However,
TOA is known for synchronization problems that can produce
estimation errors. The error in the estimation of the intervehicle
distances was simulated by a zero-mean white noise, with
the standard deviation changed from 0.5 to 10 m in half-
meter increments. For every value of the standard deviation,
the experiment was repeated ten times, and the localization
error was averaged over the 100 vehicles for the open-area
and multipath environments. Fig. 18 shows that the average
localization error of IVCAL in the multipath environment
increases with the increase in noise in the relative distance
estimation. However, the localization error of IVCAL is less
than that of the GPS/INS fusion in multipath environments.
The standard deviation of the localization error in multipath
environments using IVCAL is smaller than that using the
GPS/INS fusion in open-sky environments. In open-sky envi-
ronments, IVCAL extremely improved the localization process,
because the technique avoids the KF memory effect depicted
in Fig. 13.
690 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010
VI. CONCLUSION
Multipath conditions constitute a major cause of GPS-based
localization errors. This paper has presented a vehicle local-
ization technique that takes advantage of intervehicle commu-
nication to enhance vehicle localization accuracy. Intervehicle
communication has been used to estimate the relative locations
of vehicles in the network. Vehicle localization is then ap-
proached as a problem of fusing GPS, INS, and relative location
information obtained from multiple vehicles.
This paper has proposed a scheme to detect the presence
of multipath conditions and to mitigate their effect on vehicle
location estimation using the proposed localization technique.
To achieve maximum accuracy, vehicles exchange their loca-
tion information, along with location uncertainty indices. The
localization unit of each vehicle employs a KF to fuse its
location data with that from neighboring vehicles that satisfy
certain location accuracy criteria.
Comparative experiments have been conducted to demon-
strate improved accuracy as a result of the proposed technique.
Enhanced localization performance is reported in both open-sky
and urban canyon environments.
The lack of adequate location anchors and/or prolonged
multipath conditions represents outstanding challenges that
continue to degrade the localization accuracy performance of
the proposed technique. Map-matching localization methods
are tentative approaches that can be complementary techniques
to IVCAL in such situations. Further research work is needed
to address these issues.
REFERENCES
[1] A. Capone, I. Filippini, L. Fratta, and L. Pizziniaco, Receiver-oriented
trajectory-based forwarding, in Wireless Systems and Network Archi-
tectures in Next-Generation Internet. New York: Springer-Verlag, 2006,
pp. 2133.
[2] S. De, A. Caruso, T. Chaira, and S. Chessa, Bounds on hop distance in
greedy routing approach in wireless ad hoc networks, Int. J. Wireless
Mobile Comput., vol. 1, no. 2, pp. 131140, Feb. 2006.
[3] J. Hightower and G. Borriello, Location systems for ubiquitous comput-
ing, Computer, vol. 34, no. 8, pp. 5766, Aug. 2001.
[4] S. Vaqar and O. Basir, Trafc pattern detection in a partially deployed
vehicular ad hoc network of vehicles, IEEE Trans. Wireless Commun.,
vol. 16, no. 6, pp. 4046, Dec. 2009.
[5] A. Rae and O. Basir, A framework for visual position estimation for
motor vehicles, in Proc. 4th Workshop Positioning, Navigat. Commun.,
Mar. 2007, pp. 223228.
[6] H.-S. Tan and J. Huang, DGPS-based vehicle-to-vehicle cooperative
collision warning: Engineering feasibility viewpoints, IEEE Trans. Intell.
Transp. Syst., vol. 7, no. 4, pp. 415428, Dec. 2006.
[7] T.-S. Dao, K. Leung, C. Clark, and J. Huissoon, Markov-based lane po-
sitioning using intervehicle communication, IEEE Trans. Intell. Transp.
Syst., vol. 8, no. 4, pp. 641650, Dec. 2007.
[8] M. Sarvi and M. Kuwahara, Using ITS to improve the capacity of free-
way merging sections by transferring freight vehicles, IEEE Trans. Intell.
Transp. Syst., vol. 9, no. 4, pp. 580588, Dec. 2008.
[9] I. Skog and P. Handel, In-car positioning and navigation technologies:
A survey, IEEE Trans. Intell. Transp. Syst., vol. 10, no. 1, pp. 421,
Mar. 2009.
[10] M. Dikaiakos, A. Florides, T. Nadeem, and L. Iftode, Location-aware
services over vehicular ad hoc networks using car-to-car communication,
IEEE J. Sel. Areas Commun., vol. 25, no. 8, pp. 15901602, Oct. 2007.
[11] P. Zhou, T. Nadeem, P. Kang, C. Borcea, and L. Iftode, EZCab: A cab-
booking application using short-range wireless communication, in Proc.
3rd IEEE Int. Conf. Pervasive Comput. Commun., Mar. 2005, pp. 2738.
[12] Enhanced Wireless 911 Services, FCC, Washington, DC. [Online].
Available: http://www.fcc.gov/Bureaus/Wireless/News_Releases/1999/
nrwl9040.txt
[13] T. Melgard, G. Lachapelle, and H. Gehue, GPS signal availability in
an urban area receiver performance analysis, in Proc. IEEE Position
Location Navigat. Symp., 1994, pp. 487493.
[14] W. Kao, Integration of GPS and dead-reckoning navigation systems, in
Proc. Vehicle Navigat. Inf. Syst. Conf., Oct. 1991, vol. 2, pp. 635643.
[15] D. Bouvet and G. Garcia, Improving the accuracy of dynamic localiza-
tion systems using RTK GPS by identifying the GPS latency, in Proc.
IEEE Int. Conf. Robot. Autom., 2000, vol. 3, pp. 25252530.
[16] Q. Honghui and J. B. Moore, Direct Kalman ltering approach for
GPS/INS integration, IEEE Trans. Aerosp. Electron. Syst., vol. 38, no. 2,
pp. 687693, Apr. 2002.
[17] R. Sharaf, A. Noureldin, A. Osman, and N. El-Sheimy, Online INS/GPS
integration with a radial basis function neural network, IEEE Aerosp.
Electron. Syst. Mag., vol. 20, no. 3, pp. 814, Mar. 2005.
[18] S. Rezaei and R. Sengupta, Kalman-lter-based integration of DGPS and
vehicle sensors for localization, in Proc. IEEE Int. Conf. Mechatronics
Autom., Jul. 2005, vol. 1, pp. 455460.
[19] [Online]. Available: www.trimble.com, http://www.trimble.com/gps/
whygps.shtml
[20] J. L. Leva, An alternative closed-form solution to the GPS pseudo-range
equations, IEEE Trans. Aerosp. Electron. Syst., vol. 32, no. 4, pp. 1430
1439, Oct. 1996.
[21] J. Hoshen, The GPS equations and the problem of Apollonius, IEEE
Trans. Aerosp. Electron. Syst., vol. 32, no. 3, pp. 11161124, Jul. 1996.
[22] Javad Navigation Systems Inc. [Online]. Available: http://www.javad.
com/index.html?/jns/gpstutorial/Chapter3.html
[23] Y. Lee, Y. Suh, and R. Shibasaki, Ajax GIS application for GNSS avail-
ability simulation, KSCE J. Civil Eng., vol. 11, no. 6, pp. 303310,
Nov. 2007.
[24] B. Azimi-Sadjadi and P. Krishnaprasad, Approximate nonlinear ltering
and its application in navigation, Automatica, vol. 41, no. 6, pp. 945956,
Jun. 2005.
[25] W. Wang, Z. Liu, and R. Xie, Quadratic extended Kalman lter approach
for GPS/INS integration, Aerosp. Sci. Technol., vol. 10, no. 8, pp. 709
713, Dec. 2006.
[26] M. El-Najjar and P. Bonnifait, Road selection using multicriteria fusion
for the road-matching problem, IEEE Trans. Intell. Transp. Syst., vol. 8,
no. 2, pp. 279291, Jun. 2007.
[27] M. Porretta, P. Nepa, G. Manara, and F. Giannetti, Location, loca-
tion, location, IEEE Veh. Technol. Mag., vol. 3, no. 2, pp. 2029,
Jun. 2008.
[28] S. Capkun, M. Hamdi, and J. Hubaux, GPS-free positioning in mobile ad
hoc networks, Cluster Comput., vol. 5, no. 2, pp. 157167, Apr. 2002.
[29] S. Venkatraman, J. Caffery, Jr., and H.-R. You, Location using LOS range
estimation in NLOS environments, in Proc. 55th IEEE Veh. Technol.
Conf., 2002, vol. 2, pp. 856860.
[30] B. L. Le, K. Ahmed, and H. Tsuji, Mobile location estimator with NLOS
mitigation using Kalman ltering, in Proc. IEEE Wireless Commun.
Netw., Mar. 2003, vol. 3, pp. 19691973.
[31] M. Wylie and J. Holtzman, The non-line-of-sight problem in mobile lo-
cation estimation, in Proc. 5th IEEE Int. Conf. Universal Pers. Commun.,
Sep. 1996, vol. 2, pp. 827831.
[32] J. Caffery and G. Stber, Overview of radiolocation in CDMA cellular
systems, IEEE Commun. Mag., vol. 36, no. 4, pp. 3845, Apr. 1998.
[33] A. Benslimane, Localization in vehicular ad hoc networks, in Proc. Syst.
Commun., Aug. 2005, pp. 1925.
[34] M. Jabbour, V. Cherfaoui, and P. Bonnifait, Management of landmarks in
a GIS for an enhanced localization in urban areas, in Proc. IEEE Intell.
Vehicles Symp., 2006, pp. 5057.
[35] K. Hornik, M. Stinchcombe, and H. White, Multilayer feedforward net-
works are universal approximators, Neural Netw., vol. 2, no. 5, pp. 359
366, 1989.
[36] S. Himavathi, D. Anitha, and A. Muthuramalingam, Feedforward neural
network implementation in FPGA using layer multiplexing for effective
resource utilization, IEEE Trans. Neural Netw., vol. 18, no. 3, pp. 880
888, May 2007.
[37] J. Nocedal and S. J. Wright, Numerical Optimization, 2nd ed.
New York: Springer-Verlag, 2000.
[38] K. Langendoen and N. Reijers, Distributed localization in wireless sensor
networks: A quantitative comparison, Comput. Netw., vol. 43, no. 4,
pp. 499518, Nov. 2003.
[39] J. Rankin, An error model for sensor simulation GPS and differential
GPS, in Proc. Position Location Navigat. Symp., Apr. 1115, 1994,
pp. 260266.
[40] M. Cannon and G. LaChapelle, Analysis of a high-performance
C/A-code GPS receiver in kinematic mode, Navigation: J. Inst. Navig.,
vol. 39, no. 3, pp. 285300, 1992.
DRAWIL AND BASIR: INTERVEHICLE-COMMUNICATION-ASSISTED LOCALIZATION 691
Nabil Mohamed Drawil (M09) received the B.Sc.
degree in computer engineering from Al-Fateh Uni-
versity, Tripoli, Libya, in 1998 and the M.Sc. degree
in electrical and computer engineering from the Uni-
versity of Waterloo, Waterloo, ON, Canada, in 2007.
He is currently working toward the Ph.D. degree
with the Department of Electrical and Computer
Engineering, University of Waterloo.
During his M.Sc. study, his research was related
to localizing faulty sensors in sensor networks and
improving vehicular localization systems in multi-
path environments. His current research interests include the development of
integrity-driven hybrid fusion techniques for vehicle localization.
Otman Basir (M09) received the B.Sc. degree
in computer engineering from Al-Fateh Univer-
sity, Tripoli, Libya, the M.Sc. degree in electri-
cal engineering from Queens University, Kingston,
ON, Canada, and the Ph.D. degree in systems de-
sign engineering from the University of Waterloo,
Waterloo, ON.
He is currently a Professor with the Department
of Electrical and Computer Engineering, University
of Waterloo. He is also the Associate Director of the
Pattern Analysis and Machine Intelligence Labora-
tory. His research interests include intelligent transportation systems, embedded
real-time systems, sensor networks, sensor and decision fusion, and biologi-
cally inspired intelligence.