1188 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO.

3, SEPTEMBER 2012
Maritime Trafc Monitoring Based on
Vessel Detection, Tracking, State Estimation,
and Trajectory Prediction
Lokukaluge P. Perera, Paulo Oliveira, Senior Member, IEEE, and C. Guedes Soares
AbstractManeuvering vessel detection and tracking (VDT),
incorporated with state estimation and trajectory prediction, are
important tasks for vessel navigational systems (VNSs), as well
as vessel trafc monitoring and information systems (VTMISs)
to improve maritime safety and security in ocean navigation.
Although conventional VNSs and VTMISs are equipped with
maritime surveillance systems for the same purpose, intelligent
capabilities for vessel detection, tracking, state estimation, and
navigational trajectory prediction are underdeveloped. Therefore,
the integration of intelligent features into VTMISs is proposed in
this paper. The rst part of this paper is focused on detecting and
tracking of a multiple-vessel situation. An articial neural network
(ANN) is proposed as the mechanism for detecting and tracking
multiple vessels. In the second part of this paper, vessel state
estimation and navigational trajectory prediction of a single-vessel
situation are considered. An extended Kalman lter (EKF) is
proposed for the estimation of vessel states and further used for
the prediction of vessel trajectories. Finally, the proposed VTMIS
is simulated, and successful simulation results are presented in this
paper.
Index TermsExtended Kalman lter (EKF), neural networks,
ship detecting and tracking, ship navigational trajectory predic-
tion, vessel state estimation (VSE), vessel trafc monitoring and
information system (VTMIS).
I. INTRODUCTION
A. Maritime Transportation
S
AFETY and security issues have been highlighted in recent
years due to the challenges faced by various sea route
systems in maritime transportation. The European Union (EU)
Manuscript received August 25, 2011; revised December 11, 2011; accepted
January 20, 2012. Date of publication March 6, 2012; date of current version
August 28, 2012. The work of L. P. Perera was supported by the Doctoral
Fellowship of the Portuguese Foundation for Science and Technology (Fun-
dao para a Cincia e a Tecnologia) under Contract SFRH/BD/46270/2008.
Furthermore, this work contributes to the project of Methodology for ships
manoeuvrability tests with self-propelled models, which is being funded by the
Portuguese Foundation for Science and Technology (Fundao para a Cincia e
Tecnologia) under Contract PTDC/TRA/74332/2006. This paper was presented
in part at the 30th International Conference on Ocean, Offshore and Arctic
Engineering, Rotterdam, The Netherlands, June 2011. The Associate Editor for
this paper was Z.-H. Mao.
L. P. Perera and C. Guedes Soares are with the Centre for Marine Tech-
nology and Engineering, Instituto Superior Tcnico, Technical University
of Lisbon, Lisbon 1049-001, Portugal (e-mail: prasad.perera@mar.ist.utl.pt;
guedess@mar.ist.utl.pt).
P. Oliveira is with the Institute for Systems and Robotics, Instituto Superior
Tcnico, Technical University of Lisbon, Lisbon 1049-001, Portugal (e-mail:
pjcro@isr.ist.utl.pt).
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.2012.2187282
is surrounded by one of the busiest and most complex sea route
systems in the world. Over 90% of the EU external trade goes
by sea, and over 3.7 billion tons of freight per year is transferred
through the EU ports. In addition, the present passenger trafc
in the seas around the EU region is approximately 350 million
passenger journeys per year [1].
With the increased demand for maritime transportation,
safety and security issues become more important, and that led
to a number of monitoring mechanisms that have been proposed
by the European Commission (EC). A proposal for an effective
local community vessel monitoring mechanism is advanced in
the EC Directive 2002/59 [2], where the high density maritime
trafc regions should be equipped with vessel trafc monitoring
and information systems (VTMISs). The EU Directive 2002/59
is further structured in the following subdivisions to incorporate
the intelligent maritime trafc monitoring tasks and to improve
maritime safety and security capabilities:
1) development of the SafeSeaNet project that consists of
a pan-European electronic information system for ocean
navigation and cargo handling;
2) collaboration among the EU member states and the re-
spective maritime authorities to face maritime issues and
challenges;
3) coordination of maritime activities around the EU coast-
line for the vessels in distress;
4) codevelopment on the long distance automatic identica-
tion system (AIS) with the International Maritime Orga-
nization (IMO) [3] and the International Association of
Marine Aids to Navigation and Lighthouse Authorities;
5) codevelopment on the shore-based trafc monitoring in-
frastructure database system with the IMO.
Therefore, the subdivisions presented highlight the require-
ments in modern maritime transposition. Despite the consid-
erable amount of intelligent features developed for air [4] and
land transportation [5] systems to improve the navigation safety
and security, those facilities are still underdeveloped for vessel
navigation systems (VNSs) and VTMISs, where considerable
sectors of maritime transportation are still performed by human
operators.
Conventional VNSs and VTMISs are equipped with sev-
eral maritime surveillance mechanisms for the same purpose:
radar, laser, laser detection and ranging (ladar), automatic
radar plotting aid (ARPA), AIS, and long-range identica-
tion and tracking (LRIT) system. However, there are many
challenges faced by those maritime surveillance systems [6]:
1524-9050/$31.00 2012 IEEE
PERERA et al.: MARITIME TRAFFIC MONITORING BASED ON TRACKING AND TRAJECTORY PREDICTION 1189
large surveillance volume, synchronization among targets and
sensors, noisy signal propagation environment, and multitarget
observations. Therefore, further developments of the maritime
surveillance systems incorporated with intelligent features can
overcome the challenges faced in modern maritime trans-
portation as discussed previously. The proposed facilities in
this paper, for vessel detection, tracking, state estimation, and
navigation trajectory prediction, provide some solutions for
these tasks in VTMISs that will be the main contribution of
this paper.
B. Maritime Surveillance Systems
The rst experimental radar system was envisioned, around
1920, for ship collision prevention [7]. The track-while-scan
(TWS) radar systems are a modern invention that is able to
track multiple targets under various dynamic conditions [8].
The TWS radar consists of measurements of the target position
and radial velocity conditions that reduce associated ambiguity
in any cluttered environment, where the targets are tracked [9].
However, the radar signals attenuate with distance, weather,
and environmental conditions, leading to target tacking perfor-
mance degradation [10]. Therefore, according to the distance,
weather, and environmental conditions, frequent calibrations
are required for the improvements in radar systems [11].
Laser-based systems are proposed to overcome the problems
faced by the radar system, as previously outlined. The ex-
perimental results of laser measurement systems (LMS)-based
target tracking are proposed in [12], where targets are approx-
imated by clusters of data points, and detection and tracking
tasks, in close proximity, are illustrated. The experimental
results of ladar image-based target identication and tracking
method are presented in [13]. As illustrated by the same, the
Ladar image consists of a dense horizontal and vertically sparse
cloud of points that can also be used to identify stationary
obstacles and moving vessel/vehicles in ocean navigation.
The ARPA system provides accurate information on range
and bearing of nearby vessels, and the AIS is capable of giving
information on the vessel structural data, position, course,
speed, etc. The AIS can be integrated with a marine trafc sim-
ulator to improve navigation safety and security [14]. Although
the ARPA and AIS systems are developed to provide navigation
aids (i.e., detection and tacking tasks) to VNSs and VTMISs,
vessel state estimation, navigational trajectory prediction, and
aids on navigational decision making among vessels are still
underdeveloped.
The main advantages of VSE and navigational trajectory
prediction tasks can be categorized as the vessel intention of
navigation and the collision risk among vessels can be predicted
ahead of time. However, these tasks are also important tools to
LRIT systems, which must be formulated as an international
maritime security network. Each country has sufcient time
to evaluate the security risk that each vessel represents to
its coastline [2]. In addition, the national and international
maritime laws that must be obeyed by the respective vessels
can also be monitored so that the coastal security and safety
issues will not be compromised.
Fig. 1. Proposed VTMIS.
C. VTMIS
The proposed VTMIS is presented in Fig. 1; the system con-
sists of three main modules: 1) the vessel detection and tracking
(VDT) module; 2) the vessel state estimation and trajectory
prediction (VSETP) module; and 3) the intervessel communi-
cation module. The VDT module consists of an articial neural
network (ANN) that is formulated for a multi-VDT process.
However, the use of an ANN approach is proposed with further
perspectives on the vessel identication and classication tasks.
For that purpose, the pattern recognition and data classication
tasks are extensively tested, and successful results are reported
in [15] and [16].
The VSETP module consists of an extended Kalman lter
(EKF) that formulates vessel state estimation (i.e., position, ve-
locity and acceleration) and navigational trajectory prediction.
The EKF algorithm is proposed in this paper as the vessel state
estimator due to its capabilities of fusing nonlinear system kine-
matics with a given set of noisy measurements. Furthermore, a
curvilinear motion model (CMM) is selected to represent the
vessel motions, in the presence of disturbances, modeled as
white Gaussian noise. A linear position model is selected to
represent the measurements, in the presence of sensor noise,
also modeled as white Gaussian noise. Finally, the intervessel
communication module facilitates communication among the
VTMIS and the respective vessel communication modules.
The work presented in this paper represents one step of the
ongoing effort to formulate an intelligent navigation system in
1190 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 3, SEPTEMBER 2012
oceangoing vessels, as further described in [17] and [18]. The
organization of this paper is as follows. Section II contains an
overview of recent developments in target detection, tracking,
state estimation, and navigational trajectory prediction. The
vessel surveillance system is described in Section III. The VDT
process and the VSETP process are presented in Sections IV
and V, respectively. Section VI contains a detailed description
of the computational simulations. Finally, the conclusion and
future work are presented in Section VII.
II. RECENT DEVELOPMENTS
A. Target Detection and Tracking
Recent developments in detection and tracking of maneu-
vering targets are discussed in this section. Detection and
tracking of moving objects (DTMO) and simultaneous local-
ization and mapping (SLAM) are two important tasks that have
been developed to facilitate navigation systems. In general,
the DTMO helps to avoid collision situations, and the SLAM
helps to locate the present position and the orientation of the
navigation system. The DTMO is an important part in this
paper as detection and tacking of the vessels are the major con-
cern. The SLAM approach is mainly based on an environment
with stationary objects. However, the moving objects in the
SLAMenvironment are modelled as nose sources. Although the
DTMO and SLAM approaches are developed as two indepen-
dent research tasks, they can be complementary to each other in
navigation [19], where detecting and tracking of both stationary
and moving objects are the important scenarios in navigation.
The DTMO can be divided into four steps, as detailed in
recent studies [20]: 1) the scan process; 2) the data clustering
process; 3) the target classication process; and 4) the target
tracking process. The scan process consists of the instruments,
e.g., radar and laser systems, which can be used for identica-
tion of both moving and stationary targets. The scan process
results in the aggregation of data into geometrical data clusters,
where a cluster can be dened as a set of measured data points
that could belong to a single or multiple objects, which can also
be points, lines, and arcs of the respective targets.
To tackle the data clustering process, several target cluster-
ing methods have been used: statistical model identication,
geometrical methods [20], and competitive learning (i.e., ANN)
[21]. Inscribed angle variation and recursive line-tting meth-
ods for arc/circle and line detection, based on LMS data, are
proposed in [22]. The special constrains to joint or to fork
points should be considered in these geometrical methods in
data classication. Therefore, in general, those geometrical
constraints cannot always produce successful results due to the
complexity of the problem at hand.
An ANN approach that can overcome the failures in data
classication due to complex geometrical constrains is pro-
posed in this paper. Furthermore, there are several advantages
in the proposed approach [23]: An ANN can accommodate
self-adaptation, approximate any universal function, capture
nonlinear behavior, and compute the associated posterior
probabilities.
The target classication process consists of the separation of
the targets with respect to the geometrical shapes and sizes.
The main aim in the target classication process is to formu-
late a segment object correspondence. The correspondence is
mainly classied into general geometrical gures like circles or
polygons. The four classication methods have been proposed
in a recent study [19]: 1) feature to feature; 2) point to feature;
3) point to point; and 4) combination. However, the geometrical
gures and features identication is also successfully done by
ANNs, as further illustrated in [24].
An ANN approach is extensively implemented in recognition
of the statistical patterns in [25]. Experimental results on the
ANN approach associated with some challenging problems
faced by the target identication (i.e., small targets in high
clutter backgrounds) are presented in [26]. The object classi-
cation is one of the most successful applications in the ANN
approach [23] that can also be adopted for target identication.
A combined architecture of the ANN classier, the online
clustering algorithm, and the evidence accumulation module for
recognition and tracking of radar emitters for electronic support
system is proposed in [27]. Although the study is limited to
identifying separate clusters of data points of the respective
targets, the similar implementation of ANN can also be used
in VTMISs and VNSs to recognize other vessel and obstacles
in maritime transportation.
The target tracking process under its size constrains is con-
sidered by several studies in the recent literature. An ANN ap-
proach for target tracking based on video images is proposed in
[28]. The ANN capabilities of tracking separate clusters of data
points are illustrated in the same. Furthermore, a competitive
neural network (CNN), where the winning neuron is trained
to approximate the target location, is considered. However, the
study is limited to a constant learning rate that can degrade the
learning process. Similarly, an image-based multitarget tacking
process, using a modied Kohonen neural network, is proposed
in [29]. However, the selection of the number of neurons
with respect to the number of targets has not been properly
addressed. Therefore, the target neuron selection is addressed
in this paper, and a proper solution is proposed.
B. State Estimation and Trajectory Prediction
Recent developments in state estimation and navigational
trajectory prediction of maneuvering targets are discussed in
this section. The main functionalities of state estimation and
navigation trajectory prediction of a target can be divided into
three sections [30]: 1) development of the target motion model
(TMM); 2) measurement model and the associated techniques
(MAT); and 3) trajectory tracking and estimation (TTE).
State estimation and navigation trajectory prediction meth-
ods usually resort to mathematical models [31]. Therefore,
the identication of an accurate TMM for the maneuvering
target is an important step in its state estimation and navigation
trajectory prediction process. In general, maneuvering TMMs
used in recent literature can be divided into three categories
considering their dimensional space: 1-D, 2-D, and 3-Dmodels.
The 3-D models are commonly used in air navigation systems,
PERERA et al.: MARITIME TRAFFIC MONITORING BASED ON TRACKING AND TRAJECTORY PREDICTION 1191
and 1-D and 2-D models are commonly used in land and
maritime navigation systems.
When a single model cannot capture the required behavior
of a target, a multiple model approach has also been proposed
[32]. A multiple model approach, a constant velocity model,
and a constant speed turn model, incorporated with an un-
scented Kalman lter (KF) for curvilinear motion of a vehicle
tracking system, can be found in [33]. An image intensity-
based multiple model adaptive estimation method for tracking a
dynamic airborne vehicle is proposed in [34]. A simulation of a
two-engine aircraft that is formulated as a three-hot-spot target
in 2-D is considered in the same work.
When the TMM is not able to capture the target behavior
in maneuvering, the target tracking lter performances can be
degraded. This is mainly due to the adaptive limitations in
the target tracking lter. This situation is also categorized as
the model lter (plant lter) mismatch condition. As an exam-
ple, a fourth-order KF based on the Jerk model is presented
in [35] for the situation where the liner tracking lter (i.e.,
general KF) failed. Hence, the proper selection of the TMM
is an important part of the state estimation and navigational
trajectory prediction process. The main objective of MAT is to
develop a measurement model and its associated techniques to
observe the target present state. In general, the MAT observes
the target states in Cartesian coordinates and the measurements
in sensor coordinates (i.e., polar or spherical formations). How-
ever, varieties of coordinate systems that consists in nonlinear
combination of Cartesian sensor coordinates have also been
developed in recent studies [36].
The TTE process consists of an effective tracking algorithm
for state estimation and navigational trajectory prediction of
a maneuvering target. The target state estimation is generally
based on noisy observations that are captured by the measure-
ment sensors. However, the estimation accuracy of the target
states and the prediction accuracy of a navigational trajectory
depend on the adaptive capabilities of the estimation lter, as
well as the goodness of the measurements. The development
of the TTE can elaborate into several types of KFs in recent
studies [30].
The laser-based obstacle motion tracking in a 2-D dynamic
unconstrained environment, with the KF algorithm to predict an
obstacle motion, is presented in [37]. A target tracking method,
in combination with a particle lter and a KF using radar infor-
mation [38], and the neural KF [39] and the neural-network-
aided KF [40] have also been illustrated in the respective
literature. However, these studies assume accurate measure-
ments of the target position due to the negligible measurement
sensor noise. Target tracking based on and lters
is illustrated in [41]. However, the target tracking and state
estimation process by the lter can be limited due to a
nonuniform data rate or a nonstationary measurement noise.
Therefore, a two-stage lter is also proposed in [42].
C. Mixed Applications
Several partially integrated industrial and research applica-
tions of target detection, tracking, state estimation, and naviga-
tional trajectory prediction can also be observed in the recent
Fig. 2. Vessel surveillance system.
literature. Laser-based systems are the most popular applica-
tions due to their capabilities of target detection and tracking in
close quarter navigation.
People tracking based on laser range data, which is a mul-
tihypothesis leg-tracker integrated with the KF and a constant
velocity model, is proposed in [43]. A 2-D laser-based obstacle
tracking approach in a dynamic unconstrained environment
using the KF algorithm [37] and the particle lter with prob-
abilistic data associations to predict obstacle motions [44] are
presented in the referred studies. A laser-based system for the
detection and classication of a 3-D moving object is illustrated
in [45] and [46]. Furthermore, an experimental evaluation of
a laser measurement system for people tracking in a mobile
platform is presented in [47], and classication of people by
a mobile robot is presented in [48].
The integrated distance measurement and vision systems are
successful research applications in target detection, tacking, and
classication due to their capabilities of capturing comprehen-
sive details of the targets. A combined laser- and vision-based
approach for simultaneous detection and tracking of multiple
pedestrians based on the Bayesian method is proposed in [49].
A neural fuzzy scheme-based KF algorithm for the estimation
of accelerations in maneuvering targets is presented in [8].
The study is formulated by the measurement residual, heading
change, and Doppler shift features that are extracted from the
pulsed Doppler radar.
However, the above discussed studies are limited to either
detection and/or tracking of targets or target state estimation
and/or navigational trajectory prediction; a complete study has
not been formulated to the authors knowledge. Therefore, this
paper proposes a VTMIS that is associated with the intelligent
features of vessel detection, tracking, state estimation, and
navigation trajectory prediction; that is the main contribution
in this paper.
1192 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 3, SEPTEMBER 2012
III. VESSEL SURVEILLANCE SYSTEM
A. Introduction
A multivessel navigation situation under a vessel surveil-
lance system (i.e., radar/laser) is presented in Fig. 2. The
measurement sensor is located in the position O (0, 0), and the
ith vessel at the kth time instant is located at the position A
i
(k).
As presented in the gure, each vessel is identied as a cluster
of data points that is observed by the sensor, where each data
point is denoted by (see Fig. 2).
The sensor measurement range has been divided into several
sectors to initiate different processes (i.e., the VDT process and
the VSETP process). The sensor measurement range, i.e., the
rst circular region, is denoted by the radius R
L2
in Fig. 2. In
the region between the range R
L2
and R
L1
, the proposed VDT
process is initiated. In the region covered by the range R
L1
, the
proposed VSETP processes are executed.
B. Coordinate Transformation
After the sensor scan, the respective range r(k) R
R
and
bearing (k) R
R
values in polar coordinates are computed.
Therefore, the accumulated data vectors of range and bearing
that represent all vessels in the sensor range at each kth time
instant can be written as
r(k) = [r
1
(k) r
2
(k) r
R
(k)]
(k) = [
1
(k)
2
(k)
R
(k)] . (1)
Hence, the range and bearing values in (1) are transformed
into Cartesian coordinates in the jth position at the kth time
instant and can be written as
C
x
j
(k)=r
j
(k) cos (
j
(k)) ,
C
y
j
(k)=r
j
(k) sin(
j
(k)) . (2)
Therefore, the jth position data point of the accumulated
data cluster at the kth time instant consists of x and y coordi-
nates that are represented by [
C
x
j
(k)
C
y
j
(k)]. However, each
position data point should be normalized with respect to the
maximum range of the sensor. The data point normalization is
proposed as a requirement, where each position vector should
be normalized for the ANN, as further described in Section IV.
The normalized jth position coordinates of the vessel at the kth
time instant can be written as
x
j
(k) =
C
x
j
(k)/R
L2
; y
j
(k) =
C
y
j
(k)/R
L2
(3)
where R
L2
is the maximum range (see Fig. 2) of the mea-
surement sensor, and [x
j
(k) y
j
(k)] are the normalized x and
y coordinates, respectively. Furthermore, to complete the nor-
malization process (i.e., each position vector should be a unit
vector), the third dimension z
j
(k) is also introduced. The third
dimension is constrained as
_
x
2
j
(k) +y
2
j
(k) +z
2
j
(k) = 1. (4)
Hence, the coordinate z
j
(k) can be calculated considering
(4), which gives unit magnitude conditions for each data point
in 3-D space. Hence, the coordinate z
i
(k) can be calculated as
z
j
(k) =
_
1 x
2
j
(k) y
2
j
(k). (5)
One should note that the origin of a third dimension can be
evaluated as a transformation of 2-D position coordinates into
3-D position coordinates. Hence, the jth data point at the kth
time instant is represented by the vector p
j
(k) R
3
and can be
written as
p
j
(k) [ x
j
(k) y
j
(k) z
j
(k) ] . (6)
C. Accumulated Measurement Vector
The complete data cluster that is generated due to the vessels
in the sensor measurement range can be formulated as an ac-
cumulation of the position vectors; therefore, the accumulated
measurement vector p(k) R
R3
at the kth time instant can be
presented as (see Fig. 2)
p(k) [ p
1
(k) p
2
(k) p
R
(k) ] . (7)
The accumulated measurement vector p(k) transfers data
from the sensor unit to the VDT module for further analysis.
IV. VESSEL DETECTION AND TRACKING PROCESS
A. ANN
The theoretical foundation of articial neurons is inspired
by biological phenomena as well as the behavior of the hu-
man brain and nervous system. An articial neuron consists
of several inputs that correspond to the synapses of a bio-
logical neuron. Furthermore, it consists of one output that
corresponds to the axon of a biological neuron. Moreover,
each input is associated with a weight value that inuences the
corresponding signal over the neuron output. This concept is
mathematically formulated as a transfer function. The transfer
function calculates a sum of the net input with respect to the
neurons weight values; then, the result will be compared with
a certain threshold value to generate the neuron output [50]. The
accumulation of several neurons in series and/or parallel can be
categorized as an ANN.
B. CNN
An ANN approach has been successfully applied in the data
clustering and classication problems with respect to stationary
data conditions in recent literature [21]. However, in this paper,
a moving data cluster condition is introduced, which can be
considered as one of the novel contributions of this paper. One
of the ANNapproaches for CNN[51] associated with a learning
algorithm (i.e., modied Instar Rule) is proposed in this paper
for the detection and tracking of a multivessel situation.
The proposed CNN structure for the VDT process is pre-
sented in Fig. 3. The sensor measurement data (i.e., the ac-
cumulated measurement vector) are the input to the CNN. As
presented in Fig. 3, the CNN consists of three subunits: 1) con-
strained weight vector subunit W; 2) constrained competition
subunit C; and 3) feedback loop (i.e., Instar Rule).
PERERA et al.: MARITIME TRAFFIC MONITORING BASED ON TRACKING AND TRAJECTORY PREDICTION 1193
Fig. 3. ANN-based detection and tracking.
C. Detection and Tracking of a Multivessel Situation
The CNN is formulated for detecting and tracking vessels,
as described previously. However, each vessel is approximated
for a cluster of data points, where the CNN is trained to detect
each moving data cluster that is entering into the range R
L2
by a competition of its neurons. The vessels entering into the
range R
L2
are detected by the neurons that are assigned on the
boundary layer of the range R
L2
, as depicted in Fig. 2.
Initially, 12 neurons around the boundary layer of the range
R
L2
are assigned to detect new vessels as multiple data clusters
that are entering into the sensor range. However, the number
of neurons in the boundary layer of the range R
L2
should be
formulated with respect to the vessel trafc conditions. When a
vessel enters into the range R
L2
as a data cluster, the neuron
closest to the data cluster gets excited, and then its weight
values will adapt and will start to capture the vessel. Then, the
excited neuron will be added as a new neuron into the ANN;
the weight values of the excited neuron will reset to the original
values. Therefore, 12 neurons will always be located alone the
boundary layer of the range R
L2
to detect new vessel entering
into the sensor range.
In this approach, the number of neurons will always increase
with the number of vessels entering the range R
L2
. Therefore,
the growing number of neurons with respect to the available
computational power will always be a challenge. To overcome
this problem, the following solution is proposed: The system
monitors weight value changes of the respective neurons, and
the weight unchanged neurons other than the boundary layer
neurons will be removed from the ANN. The condition can
also be seen as the removal of a neuron due to the respective
vessel exiting from the sensor range. This proposed mechanism
is implemented on the constrained weight vector subunit (see
Fig. 3). Furthermore, the weight values of the prototype neurons
(i.e., initial neurons) W(k) R
S3
are stored as row vectors in
the weight vector subunit, which can be written as
W(k) = [w
1
(k) w
2
(k) w
S
(k)] . (8)
Furthermore, the jth weight value of the respective neu-
ron is represented by the vector w
j
(k). However, for a fair
competition among neurons, each position vector p
i
(k) and
each prototype vector w
j
(k) should have unit magnitude con-
ditions as discussed previously. Hence, the position vector
magnitude and the neuron weight magnitude should satisfy the
conditions
|p
i
(k)| =1
|w
j
(k)| =1. (9)
The accumulated net input n(k) R
S
is the output from the
constrained weight vector subunit. The accumulated net input
n(k) can be written as
n(k) = [n
1
(k) n
2
(k) n
S
(k)] . (10)
The net input of the ith neuron n
i
(k) is the scalar product
between the position vector p
i
(k) and the prototype vector (i.e.,
prototype neuron) w
j
(k). The ith net input n
i
(k), considering
the unit magnitude conditions in the position vector p
i
(k) and
the prototype vector w
j
(k), can be written as
n
i
(k) = W(k)p
i
(k) =
_

_
w
T
1
(k)p
i
(k)
w
T
2
(k)p
i
(k)
.
.
.
w
T
S
(k)p
i
(k)
_

_
=
_

_
cos
1
(k)
cos
2
(k)
.
.
.
cos
S
(k)
_

_
.
(11)
The ith net input n
i
(k) will input to the constrained competi-
tion subunit C that consists of the cosine angle values between
the position vector p
i
(k) and the prototype vector w
j
(k).
D. Neuron Competition
In the constrained competition subunit C, the distances
between each position vector p
i
(k) to each prototype vector
w
j
(k) is compared; this has been done by the mathematical
formulation described in (11). Then, the neuron whose weight
vector w
j
(k) is in the direction closest to the position vec-
tor p
i
(k) will assign an output of 1, and others will assign
0, by the transfer function compet(.). This transfer function
is formulated in such a way as to generate fair competition
among neurons. Therefore, the proposed competition can be
written as
a
i
(k) = compet (n
i
(k)) = compet (W(k)p
i
(k)) (12)
where a
j
(k) is the output from the transfer function. The
competitive transfer function can be illustrated as
compet (n
i
(k)) =
_
1, for neuron with max n
i
(k)
0, all other neurons.
The accumulated net input n(k) and the accumulated net
output a(k) R
S
are the associated constrained competition
subunit C at the kth time instant. Finally, the feedback loop
associated with the Instar Rule is proposed to adapt the weight
values of the prototype vector considering the accumulated net
output a(k).
At the initial stage, the weight values of the prototype
neurons of the CNN are assumed to be known and located
around the boundary layer of the range R
L2
. However, to
track the moving vessels that are represented by the clusters
of data points, new neurons are introduced, and their weight
values will adapt, resorting to a learning rule. The unsupervised
1194 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 3, SEPTEMBER 2012
learning rule proposed in this section works as described next:
When a competitive layer excites a neuron that is closest to
the data cluster, then the learning rule will adapt appropriate
weight values in the prototype neuron that moves closer to the
respective data cluster. The Instar Rule is considered in this
paper as an unsupervised learning approach.
E. Instar Rule
The Instar Rule is derived from the Hebb Rule, as illustrated
in [51]. The unsupervised Hebb Rule can be written as
W(k) = W(k 1) +a
i
(k)p
T
i
(k) (13)
where is the learning rate. Note that a constant learning rate
is a disadvantage in the training process of the CNN. Thus,
the rate of error convergence must be altered. Although at the
beginning of the training process a higher learning rate has
advantages to the CNN regarding the error minimization, a
high learning rate will present disadvantages in terms of error.
Hence, to improve the Hebb Rule, a weight decaying term that
is proportional to a
i
(k)W(k 1) is introduced. The modied
Hebb Rule can then be written as
W(k) = W(k 1) +a
i
(k)p
T
i
(k) a
i
(k)W(k 1) (14)
where is the decay rate. Assuming = , (14) can further be
illustrated as
W(k) = W(k 1) +a
i
(k)
_
p
T
i
(k) W
i
(k 1)
_
. (15)
This is the Instar Rule that is proposed as the unsupervised
learning rule for training of the CNN in this paper.
F. Constrained Neural Competition
The neural competition proposed in this paper can be sum-
marized as a situation where the closest neuron gets excited by
a data cluster and then the excited neuron will take over all
the data points in the respective data cluster. However, after
winning the data cluster, the weight values of the respective
neuron should be updated. Therefore, the wining neuron is
updated and moved closer to the respective data cluster. This
process will continue with the moving data clusters that are
represented by the target vessels at each time instant.
However, several drawbacks can be observed in this ap-
proach. When the CNN is under complex environmental con-
ditions, with multiple stationary and moving targets, several
neurons can track different parts of the same target, and/or one
neuron can track several targets in close proximity conditions
[12]. As a solution to this problem, several constrains are
introduced into the constrained weight vector subunit W and
the constrained competition subunit C.
The rst constrain is introduced in the region between the
range R
L1
and R
L2
. When a target vessel enters this region, the
neuron in the boundary layer, which is closer to the vessel, gets
excited and starts to tack the respective data cluster. However,
in some situations, two neurons can be excited under the same
conditions. A situation of two close neurons, excitement can
happen either as two vessels are entering the region or a single
vessel is entering between those neurons. Therefore, to over-
come this drawback, a constraint condition is introduced into
the constrained weight vector subunit W. When two neurons
in close proximity are excited, a condition to check the distance
between two neurons is introduced; if the distance between two
neurons is smaller than a given threshold, then both neurons
are considered as a single neuron. Therefore, this corresponds
to an event, where a single vessel is entering into the sensor
region. Otherwise, the situation is considered as two vessels are
entering into the sensor region.
The second constrain is introduced in the range R
L1
. When
several data clusters get closer to each other, the respective
neurons of the data clusters can jump among other data clusters
in close proximity. This is similar to a situation where several
neurons are tracking different parts of the same target and/or a
single neuron is tracking several targets as discussed previously.
Therefore, to overcome these types of failures, a constrain into
the constrained competition subunit C is introduced.
When a neuron gets excited by a data point, a condition to
check the distance between the neuron and the data point is
introduced. If the distance between the neuron and the data
point is greater than a given threshold, then the data cluster will
be removed from the respective neuron competition. Therefore,
this mechanism limits neuron movements among data clusters
in close proximity and facilitates a competition inside the
respective data cluster. The properly dened constrains on the
CNN can reduce these ill effects from the neural competition.
Finally, the CNN will formulate a mean position value for each
tracking vessel. The mean position value of each tracking vessel
will be transferred to the VSETP module for further analysis.
V. STATE ESTIMATION AND TRAJECTORY PREDICTION
A. Introduction
The estimated mean position of each tracking vessel by the
CNN is further analyzed in this section to estimate vessel states
and to predict navigational trajectories. Therefore, this section
elaborates on three divisions: 1) TMM; 2) MAT; and 3) TTE.
B. TMM
A suitable mathematical model for tacking an ocean going
vessel is considered in this section. A 2-D kinematic model that
can capture the vessel maneuvering behavior is considered. In
general, an ocean going vessel always follows slow parabolic-
type maneuvers, where the fast-changing maneuvers are ob-
served in land and airborne vehicles. The TMM is assumed to
be a point object with negligible dimensions corresponding to
the mean vessel position that is estimated by the CNN. Consid-
ering the discussed requirements, a continuous-time CMM is
proposed as the TMM.
The continuous-time CMM that is formulated for vessel
navigation is presented in Fig. 4. The vessel is located at
PERERA et al.: MARITIME TRAFFIC MONITORING BASED ON TRACKING AND TRAJECTORY PREDICTION 1195
Fig. 4. CMM.
point A. The vessel position coordinates are represented by
x(t) and y(t) in continuous time. Furthermore, the continuous-
time velocity components along the x- and y-directions are
represented by v
x
(t) and v
y
(t). The course of the vessel is
presented by
a
(t). It is also assumed that the vessel course and
heading conditions are similar. The vessel speed is presented
by V
a
(t), as illustrated in the gure. One should note that
the actual vessel orientation cannot be observed by the sensor
measurements.
Some important features can be observed from the CMM:
as presented in Fig. 4, under the null normal acceleration,
a
n
(t) = 0, the model performs straight line motions; and when
tangential acceleration, a
t
(t) = 0, the model performs circu-
lar motions. Furthermore, the different acceleration conditions
a
t
(t) > 0 and a
t
(t) < 0 produce various parabolic navigation
trajectories, as presented in the gure. Therefore, the CMM
capabilities of capturing the multimodel features are the main
advantages in this approach. A standard continuous-time CMM
can be written as [52]

a
(t) =
a
n
(t)
V
a
(t)
;

V
a
(t) = a
t
(t)
v
x
(t) =V
a
(t) sin (
a
(t)) ; v
y
(t) = V
a
(t) cos (
a
(t)) . (16)
The summarized CMM, considering (16), can be formulated
as a nonlinear TMM. The TMM can be written as
x
x
(t) = f (x
x
(t)) +w
x
(t) (17)
where the system states x
x
(t) and the system function f(x
x
(t))
can further be illustrated as
x
x
(t) =
_

_
x(t)
v
x
(t)
y(t)
v
y
(t)
a
t
(t)
a
n
(t)
_

_
, f (x
x
(t)) =
_

_
v
x
(t)
a
t
(t)f
vx
+a
n
(t)f
vy
v
y
(t)
a
t
(t)f
vy
a
n
(t)f
vx
0
0
_

_
f
vx
=
v
x
(t)
_
v
2
x
(t) +v
2
y
(t)
, f
vy
=
v
y
(t)
_
v
2
x
(t) +v
2
y
(t)
and w
x
(t) is a white Gaussian process noise with 0 mean and
Q(t) covariance that can be written as (18), shown at the bottom
of the page, where Q
x
(t), Q
vx
(t), Q
y
(t), Q
vy
(t), Q
at
(t), and
Q
an
(t) are the respective state covariance values. Furthermore,
the tangential acceleration a
t
(t) and the normal acceleration
a
n
(t) derivatives can be written as
a
t
(t) =w
at
(t); a
t
(t
0
) = a
t0
a
n
(t) =w
nt
(t); a
n
(t
0
) = a
n0
where a
t0
and a
n0
are the initial acceleration values, w
at
(t) and
w
nt
(t) are the derivates of the tangential and normal accelera-
tions, which are modeled as white Gaussian distributions with
zero mean and Q
at
(t) and Q
an
(t) covariances, respectively.
The Jacobian of the system function f(x
x
(k)) can be expressed
as (19), shown at the bottom of the page, where the respective
functions in (19) can be written as
f
vx
vx
=
v
2
y
(t)
_
v
2
x
(t) +v
2
y
(t)
_
3/2
, f
vx
vy
=
v
y
(t)v
x
(t)
_
v
2
x
(t) +v
2
y
(t)
_
3/2
f
vy
vx
=
v
x
(t)v
y
(t)
_
v
2
x
(t) +v
2
y
(t)
_
3/2
, f
vy
vy
=
v
2
x
(t)
_
v
2
x
(t) +v
2
y
(t)
_
3/2
.
C. MAT
The measurement model is formulated as a discrete-time
linear model due to the availability of the ocean vessel position
values in discrete time instants. The position measurement
sensor is located in the origin O (0, 0), as presented in
Fig. 1. It is assumed that the correlations between the position
Q(t) = diag [ Q
x
(t) Q
vx
(t) Q
y
(t) Q
vy
(t) Q
at
(t) Q
an
(t) ] (18)

x
x
(f (x
x
(t))) =
_

_
0 1 0 0 0 0
0 a
t
(t)f
vx
vx
+a
n
(t)f
vy
vx
0 a
t
(t)f
vx
vy
+a
n
(t)f
vy
vy
f
vx
f
vy
0 0 0 1 0 0
0 a
t
(t)f
vy
vx
a
n
(t)f
vx
vx
0 a
t
(t)f
vy
vy
a
n
(t)f
vx
vy
f
vy
f
vx
0 0 0 0 0 0
0 0 0 0 0 0
_

_
(19)
1196 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 3, SEPTEMBER 2012
measurements are negligible. The vessel position measurement
model in discrete time can be written as
z
z
(k) = h(x
x
(k)) +w
z
(k) (20)
where the system states z
z
(k) and the measurement function
h(x
x
(k)) can further be illustrated as
z
z
(k) =
_
z
x
(k)
z
y
(k)
_
h(x
x
(k)) =
_
x(k) 0 0 0 0 0
0 0 y(k) 0 0 0
_
where z
x
(k) and z
y
(k) are the measurements of x and y
position values of the vessel, and w
z
(k) is white Gaussian
measurement noise with zero mean and R(k) covariance. The
covariance R(k) can be written as
R(k) = diag [ R
x
(k) R
y
(k) ] . (21)
The Jacobian matrix of the measurement function h(x
x
(k))
can be written as

x
x
(h(x
x
(k))) =
_
1 0 0 0 0 0
0 0 1 0 0 0
_
. (22)
D. TTE
In 1960, Kalman solved the problem of minimizing the
mean least square error for linear systems under Gaussian
disturbances [52], i.e., a ltering problem, using the state space
approach. This solution is now denominated as the KF. The
proposed approach is formulated with two main features: 1) the
vector modeling the random processes, and 2) the process
noise are Gaussian [53]. However, the general KF algorithm
is limited for application of linear systems. Therefore, the EKF
is considered as a standard technique for a number of nonlinear
system applications. The EKF algorithm can be summarized as
follows [54]:
1) System model
x
x
(t) = f (x
x
(t)) +w
x
(t); E [w
x
(t)] = 0
E [w
x
(t); w
x
(t)] = [Q(t)] (23)
2) Measurement model
z
z
(k) =h(x
x
(k)) +w
z
(k), k = 1, 2, . . .
E [w
z
(k)] =0, E [w
z
(k); w
z
(k)] = [R(k)] . (24)
3) Error conditions
x
x
(k) = x
x
(k) x
x
(k) (25)
where x
x
(t) is the state error vector, and x
x
(t) is the
estimated states of the system.
4) System initial states
x
x
(0) N ( x
x
(0), P(0)) (26)
where P(0) is the state initial covariance, describing the
uncertainty present on the initial estimates.
5) Other conditions
E [w
x
(t); w
z
(k)] = 0 for all k, t. (27)
6) State estimation propagation
d
dt
x
x
(k) = f ( x
x
(k)) . (28)
7) Error covariance propagation
d
dt
P(k) =F ( x
x
(k)) P(k) +P(k)F
T
( x
x
(k)) +Q(k)
F ( x
x
(k)) =

x
x
(k)
f (x
x
(k))

x
x
(k)= x
x
(k)
(29)
where P(k) is the estimated error covariance that can be
written as (30), shown at the bottom of the page, and
P
x
(k), P
vx
(k), P
y
(k), P
vy
(k), P
at
(k), and P
an
(k) are
the respective estimated state error covariance values.
8) State estimate update
At each step, after the measurement data are available
from the CNN, the system state update can be esti-
mated as
x
x
(k
+
) = x
x
(k

) +K(k)
_
z
z
(k) h
_
x
x
(k

)
_
(31)
where x
x
(k

) and x
x
(k
+
) are the prior and posterior
estimated system states, respectively, and K(k) is the
Kalman gain.
9) Error covariance update
P(k
+
) =
_
1 K(k)H
_
x
x
_
k

__
P(k

) (32)
where P(k

) and P(k
+
) are the prior and posterior error
covariances of the system, respectively, and
H ( x
x
(k)) =

x
x
(k)
h(x
x
(k))

x
x
(k)= x
x
(k)
. (33)
10) KF gain
K(k) = P(k

)H
_
x
x
(k

)
_

_
H
_
x
x
(k

)
_
P(k

)H
_
x
x
(k

)
_
T
+R(k)
_
1
. (34)
VI. COMPUTATIONAL SIMULATIONS
The computational simulations of vessel detection, tracking,
state estimation, and navigational trajectory prediction based
on a multiple vessel situation are presented in this section.
P(k) = diag [ P
x
(k) P
vx
(k) P
y
(k) P
vy
(k) P
at
(k) P
an
(k) ] (30)
PERERA et al.: MARITIME TRAFFIC MONITORING BASED ON TRACKING AND TRAJECTORY PREDICTION 1197
Fig. 5. VDT.
The proposed systems are simulated on the MATLAB software
platform. The system consists of three main computational
loops: 1) target scanning loop; 2) VDT loop; and 3) VSETP
loop. The main objective of the target scanning loop is to scan
the environment and to observe the vessels as data clusters.
Then, these data will be transferred into the VDT module. The
main objective of the VDT module is to adapt the CNN to
track the positions of the respective vessels by updating the
neuron weights by the learning rule. Finally, the vessels mean
positions will be transferred into the VSETP module. The main
objective of the VSETP module is to use the EKF algorithm to
estimate respective vessel states and to predict its navigational
trajectory.
A. VDT
The computational simulations of tracking a three-vessel
situation, represented by three groups of data clusters, are
presented in Fig. 5. The three data clusters are moving under
varying acceleration conditions with actual trajectories, respec-
tively. As depicted in the gure, the initial weight values of the
CNN are assigned along the boundary layer of the range R
L2
.
As presented in the gure, the respective neurons adapt their
weight values to track the movements of each cluster. Further-
more, it can be observed that each tracking neuron converges
to an approximate mean value of the respective data cluster.
This mean value of a vessel is considered as the measurement
position at each time instant. The vessel mean position values
are forwarded to the VSETP module. As presented also in the
gure, the tracking neurons are shown as small circles. The
estimated vessel position values are presented by the center of
the respective small circles.
B. VSETP
The rst set of computational simulations for VSETP pre-
diction based on the EKF algorithm is presented in Figs. 68.
Fig. 6. Navigational trajectories: Situation 1.
Fig. 7. Vessel velocity conditions: Situation 1.
The vessel navigation conditions under a circular maneuver are
considered in Fig. 6. The gure presents the actual circular
trajectory (Act. Traj), the measured trajectory (Mea. Traj.),
and the EKF estimated trajectory (Est. Traj.) of vessel nav-
igation. The measured trajectory data have been simulated
by adding sensor noise into the actual trajectory data. As
noted from the gure, the actual vessel trajectory is estimated
successfully by the EKF. The vessel velocity components of
x-direction v
x
(t) and y-direction v
y
(t) of the actual and esti-
mated values are presented in Fig. 7. As noted from the gure,
the vessel velocity components are also estimated successfully
by the EKF.
The tangential acceleration a
t
(t) and the normal acceleration
a
n
(t) of the actual and estimated values are presented in Fig. 8.
As noted from the gure, the vessel acceleration components
are also estimated successfully by the EKF. Furthermore, these
acceleration conditions can be used for the prediction of future
navigation trajectories, which is another contribution in this
paper.
1198 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 3, SEPTEMBER 2012
Fig. 8. Vessel acceleration conditions: Situation 1.
Fig. 9. Navigational trajectories: Situation 2.
The second set of computational simulations for VSETP
based on the EKF algorithm is presented in Figs. 911. The
vessel navigation conditions of an approximately straight line
maneuver are considered in this situation. Fig. 9 presents the ac-
tual trajectory alone with the measured trajectory and the EKF
estimated trajectory of the vessel. As noted from the gure,
the EKF estimates the vessel navigation trajectory successfully.
The vessel velocity components of the x-direction v
x
(t) and
y-direction v
y
(t) of the actual and estimated values are pre-
sented in Fig. 10. As noted in the gure, the vessel velocity
components are also estimated successfully by the EKF. The
tangential acceleration a
t
(t) and the normal acceleration a
n
(t)
of the actual and estimated values are presented in Fig. 11. As
noted in the gure, the vessel acceleration components are also
estimated successfully by the EKF. As mentioned before, these
acceleration conditions can be used for the prediction of future
navigation trajectory.
VII. CONCLUSION AND FUTURE DEVELOPMENT
The novelty of this paper is that the target dimensions are
explicitly considered during the detection and tracking process,
departing from the simplistic assumptions that most of the
target tracking methods assume, i.e., the target is one point or
Fig. 10. Vessel velocity conditions: Situation 2.
Fig. 11. Vessel acceleration conditions: Situation 2.
approximated by a small data cluster. Furthermore, a popular
state-of-the-art machine learning application for an ANN asso-
ciated with unsupervised learning is successfully implemented
and validated in this paper. Although ANN applications are ex-
tensively used for the recognition of clusters of static data pat-
terns, the tracking of clusters of moving data patterns can also
be successfully implemented. Furthermore, the ANN approach
can further be developed not only for vessel tracking but for
the tasks of classication and identication of vessels as well.
Hence, advanced ANN approaches for vessel classication and
identication will be considered as future developments of this
paper.
The data association and occlusion conditions are some of
the challenges faced by the ANN approach. As presented in
Section VI, adding constrains into the ANNcan overcome some
of those issues. Furthermore, the AIS and LRIT systems made
use of a considerable number of ocean going vessels, where the
vessels identication information as well as the vessels course
and speed conditions are transmitted by individual vessels.
Therefore, the proposed VTMIS can be integrated with the AIS
and LRIT system information, which is another solution to the
data association and occlusion problems.
The VSE (i.e., position, velocity, and acceleration) process
has successfully been achieved by the EKF, given that a CMM
was selected, under a linear measurement model provided by
PERERA et al.: MARITIME TRAFFIC MONITORING BASED ON TRACKING AND TRAJECTORY PREDICTION 1199
commercially available ranging sensors. The estimated values
of velocity and acceleration components have small errors,
varying around the actual values with negligible errors. The
vessel navigation consists of changing acceleration conditions,
as presented in this paper. One should note that the accel-
eration estimations are achieved only by using the position
measurements that are estimated by the ANN. Furthermore,
the estimated vessel states can be used for the estimation of
future navigation trajectories, which eventually help to predict
the future behavior, the intention, and the respective collision
risk among vessels ahead of time.
REFERENCES
[1] Operational Activities: European Maritime Safety Agency, EMSA Rep.,
Jan. 2009. [Online]. Available: http://www.emsa.europa.eu/
[2] Directive 2002/59/EC of the European Parliament and of the Council,
Establish a community vessel trafc monitoring and information system
and repealing council directive 93/75/EEC, Ofcial J. Eur. Commun.,
no. 208, pp. 10272002.
[3] Int. Maritime Org. [Online]. Available: http://www.IMO.org
[4] M. Prandini, L. Piroddi, S. Puechmorel, and S. L. Brzdilov, Toward air
trafc complexity assessment in new generation air trafc management
systems, IEEE Trans. Intell. Transp. Syst., vol. 12, no. 3, pp. 809818,
Sep. 2011, DOI: 10.1109/TITS.2011.2113175.
[5] U. Franke and S. Heinrich, Fast obstacle detection for urban trafc
situations, IEEE Trans. Intell. Transp. Syst., vol. 3, no. 3, pp. 186197,
Sep. 2002.
[6] D. L. Hall and J. Llinas, An introduction to multisensor data fusion,
Proc. IEEE, vol. 85, no. 1, pp. 623, Jan. 1997.
[7] F. Farina, Introduction to radar signal & data processing: The opportu-
nity, in RTO SET Lecture Series on: Knowledge-Based Radar Signal and
Data Processing. Neuilly-sur-Seine Cedex, France: RTO, Nov. 2004.
[8] F. Duh and C. Lin, Tracking a maneuvering target using neural fuzzy
network, IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 34, no. 1,
pp. 1633, Feb. 2004.
[9] A. Farina and S. Pardini, Track-while-scan algorithm in clutter
environment, IEEE Trans. Aerosp. Electron. Syst., vol. AES-14, no. 5,
pp. 769779, Sep. 1987.
[10] S. E. Giangrande and A. V. Ryzhkov, Calibration of dual-polarization
radar in the presence of partial beam blockage, J. Atmos. Ocean. Tech-
nol., vol. 22, no. 8, pp. 11561166, Aug. 2005.
[11] R. Naranjo, Radar Revisited: Ocean Navigator Online. [Online]. Avail-
able: http://www.oceannavigator.com/
[12] L. P. Perera and C. Guedes Soares, Laser measurement system based
maneuvering target tracking formulated by adaptive competitive neural
networks, in Proc. 2nd Int. Conf. Adapt. Self-Adapt. Syst. Appl., Lisbon,
Portugal, 2010, pp. 8490.
[13] A. R. J. Ruiz and F. S. Granja, A short-range ship navigation system
based on ladar imaging and target tracking for improved safety and ef-
ciency, IEEE Trans. Intell. Transp. Syst., vol. 10, no. 1, pp. 186197,
Mar. 2009.
[14] K. Hasegawa, Advanced marine trafc automation and management sys-
tem for congested waterways and coastal areas, in Proc. ICOE, 2009,
pp. 110.
[15] I. Kumazawa, Target tracking by matching a shape represented by a tree
of sigmoid functions, Pattern Recognit. Lett., vol. 21, no. 6/7, pp. 661
675, Jun. 2000.
[16] J. Zhao, A novel recurrent neural network for face recognition, J. Softw.,
vol. 12, no. 8, pp. 11281139, 2001.
[17] L. P. Perera, J. P. Carvalho, and C. Guedes Soares, Bayesian network
based sequential collision avoidance action execution for an ocean naviga-
tional system, in Proc. 8th IFAC Conf. Control Appl. Mar. Syst., Rostock,
Germany, 2010, pp. 301306.
[18] L. P. Perera, P. Oliveira, and C. Guedes Soares, Collaborated and con-
strained neural-EKF algorithm for the vessel trafc monitoring and in-
formation system, in Proc. 30th Int. Conf. Ocean, Offshore Arctic Eng.,
Rotterdam, The Netherlands, Jun. 2011, pp. 871879.
[19] C. Wang and C. Thorpe, Simultaneous localization with detection and
tracking of moving objects, in Proc. IEEE Int. Conf. Robot. Autom.,
Washington, DC, 2002, pp. 29182924.
[20] A. Mendes, L. C. Bento, and U. Nunes, Multi-target detection and track-
ing with a lasers canner, in Proc. IEEE Intell. Veh. Symp., Parma, Italy,
2004, pp. 796800.
[21] K. L. Du, Cluster: A neural network approach, Neural Netw., vol. 23,
no. 1, pp. 89107, Jan. 2010.
[22] J. Xavier, M. Pacheco, D. Castro, A. Ruano, and U. Nunes, Fast line
arc/circle and leg detection from Laser scan data in a player driver, in
Proc. IEEE Int. Conf. Robot. Autom., Apr. 2005, pp. 39303935.
[23] G. P. Zhang, Neural networks for classication: A survey, IEEE Trans.
Syst., Man, Cybern. C, Appl. Rev., vol. 30, no. 4, pp. 451462, Nov. 2000.
[24] Z. Jie-yu, A novel recurrent neural network for face recognition, J.
Softw., vol. 12, no. 8, pp. 11281139, 2001.
[25] C. M. Bishop, Neural Networks for Pattern Recognition. Oxford, U.K.:
Clarendon, 1995.
[26] M. V. Shirvaikar and M. Trivedi, A neural network lter to detect small
targets in high clutter backgrounds, IEEE Trans. Neural Netw., vol. 8,
no. 1, pp. 252255, Jan. 1995.
[27] E. Granger, M. A. Rubin, S. Grossberg, and P. Lavoie, A what-and-where
fusion neural network for recognition and tracking of multiple Radar
emitters, Boston Univ., Boston, MA, Tech. Rep. CAS/CNS-TR-2000-
029, Dec. 2000.
[28] E. Araujo, C. R. Silva, and D. J. B. S. Sampaio, Video target tracking
by using competitive neural networks, WSEAS Trans. Signal Process.,
vol. 4, no. 8, pp. 420431, Aug. 2008.
[29] F. Gracia and E. Arajo, Visual multi-target tracking by using modied
Kohonen neural networks, in Proc. Int. Joint Conf. Neural Netw., 2008,
pp. 41634167.
[30] X. R. Li and V. P. Jilkov, A survey of maneuvering target tracking
Part IV: Decision-based methods, in Proc. SPIE Conf. Signal Data
Process. Small Targets, Orlando, FL, 2002, pp. 47284760.
[31] X. R. Li and V. P. Jilkov, Survey of maneuvering target tracking
Part I: Dynamic models, IEEE Trans. Aerosp. Electron. Syst., vol. 39,
no. 4, pp. 13331364, Oct. 2003.
[32] M. Athans and C. B. Chang, Adaptive estimation and parameter identi-
cation using multiple model estimation algorithm, Lincoln Lab, Mass.
Inst. Technol., Lexington, MA, Tech. Note 1976-28, 1976.
[33] Y. Kim and K. Hong, An IMM algorithm for tracking maneuvering
vehicles in an adaptive cruise control environment, Int. J. Control Autom.
Syst., vol. 2, no. 3, pp. 310318, 2004.
[34] P. Maybeck, Adaptive tracking of dynamic airborne vehicles based on
(ir) image plane intensity data, in Lecture notes in Control and In-
formation Science. Berlin, Germany: Springer, 1985, ser. Stochastic
Differential Systems, pp. 284304.
[35] K. Mehrotra, A jerk model for tracking highly maneuvering targets,
IEEE Trans. Aerosp. Electron. Syst., vol. 33, no. 4, pp. 9821094,
Oct. 1997.
[36] X. R. Li and V. P. Jikov, A survey of maneuvering target tracking
Part III: Measurement models, in Proc. SPIE Conf. Signal Data Process.
Small Targets, San Diego, CA, Aug. 2001, vol. 4473, pp. 423426.
[37] M. Berker, R. Hall, S. Kolski, K. Macek, R. Siegwart, and B. Jensen, 2d
Laser-based probabilistic motion tracking in urban-like environments,
J. Brazilian Soc. Mech. Sci. Eng., vol. 31, no. 2, pp. 8396, 2009.
[38] D. Angelova and L. Mihaylova, Joint target tracking and classica-
tion with particle ltering and mixture Kalman ltering using kinematic
radar information, Digit. Signal Process., vol. 16, no. 2, pp. 180204,
Mar. 2006.
[39] S. C. Stubberud and K. A. Kramer, Kinematic prediction for intercept
using a neural Kalman lter, in Proc. IFAC World Congr., Prague, Czech
Republic, 2005.
[40] V. Vaidehi, N. Chitra, M. Chokkalingam, and C. N. Krishnan, Neural net-
work aided Kalman lter for multitarget tracking applications, Comput.
Elect. Eng., vol. 27, no. 2, pp. 217228, May 2001.
[41] P. R. Kalta, The tracking index: A generalized parameter of and
target tracking, IEEE Trans. Aerosp. Electron. Syst., vol. AES-
20, no. 2, pp. 174182, Mar. 1984.
[42] W. D. Blair, Two-stage alpha-beta-gamma estimator for tracking maneu-
vering targets, in Proc. Amer. Control Conf., Chicago, IL, Mar. 2009,
pp. 842846.
[43] K. O. Arras, S. Grzonka, M. Luber, and W. Burgard, Efcient people
tracking in Laser range data using a multi-hypothesis leg-tracker with
adaptive occlusion probabilities, in Proc. IEEE Int. Conf. Robot. Autom.,
Pasadena, CA, 2008, pp. 17101715.
[44] A. Almeida, J. Almeida, and R. Araujo, Real-time tracking of multiple
moving objects using particle lters and probabilistic data association,
Automatika, vol. 46, no. 1/2, pp. 3948, 2005.
[45] A. Lourenco, P. Freitas, M. I. Ribeiro, and J. S. Marques, Detection and
classication of 3D moving objects, in Proc. 10th IEEE Mediterranean
Conf. Control Autom., Lisbon, Portugal, 2002.
[46] G. Gallagher, S. S. Srinivasa, J. A. Bagnell, and D. Ferguson, GATMO:
A generalized approach to tracking movable objects, in Proc. IEEE Int.
Conf. Robot. Autom., 2009, pp. 43224327.
1200 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 3, SEPTEMBER 2012
[47] M. Linstrom and J. O. Eklundh, Detection and tracking moving objects
from a mobile platform using a Laser range scanner, in Proc. IEEE/RSJ
Int. Conf. Intell. Robots Syst., Maui, HI, 2001, pp. 13641369.
[48] M. Luber, K. O. Arras, C. Plagemann, and W. Burgard, Classifying
dynamic objects: An unsupervised learning approach, in Proc. Robot.,
Sci. Syst. IV, Zurich, Switzerland, 2008.
[49] X. Song, J. Chi, H. Zhao, and H. Zha, A Bayesian approach: Fusion of
laser and vision for multiple pedestrians tracking, Int. J. Adv. Comput.
Eng., vol. 3, no. 1, pp. 19, Jun. 2008.
[50] M. Cirstea, A. Dinu, J. Khor, and M. McCormick, Eds., Neural and Fuzzy
Logic Control of Drives and Power Systems, 1st ed. New York: Elsevier,
2002.
[51] M. T. Hagan, H. B. Demuth, and M. H. Beale, Eds., Neural Network
Design. Boston, MA: PWS, 1996.
[52] R. A. Best and J. P. Norton, A new model and efcient tracker fro target
with curvilinear motion, IEEE Trans. Aerosp. Electron. Syst., vol. 33,
no. 3, pp. 10301037, Jul. 1997.
[53] R. G. Brown and P. Y. C. Hwang, Introduction to Random Signals and
Applied Kalman Filtering. New York: Wiley, 1997.
[54] A. Gelb, J. F. Kasper, Jr., R. A. Nash, Jr., C. F. Price, and A. A. Sutherland,
Jr., Applied Optimal Estimation. Cambridge, MA: MIT Press, 2000.
Lokukaluge P. Perera received the B.S. and M.S.
degrees in mechanical engineering from Oklahoma
State University, Stillwater, in 1999 and 2001,
respectively.
He is currently with the Centre for Marine Tech-
nology and Engineering (CENTEC), Instituto Supe-
rior Tcnico, Technical University of Lisbon, Lisbon,
Portugal. He has won a Doctoral Fellowship from the
Portuguese Foundation for Science and Technology
and is currently conducting Doctoral studies in naval
architecture and marine engineering with the Tech-
nical University of Lisbon. At CENTEC, he is working on the use of articial
intelligence and control techniques on autonomous ocean navigation systems.
Paulo Oliveira (S91M95SM11) received the
Licenciatura, M.S., and Ph.D. degrees in electrical
and computer engineering from Instituto Superior
Tcnico (IST), Lisbon, Portugal, in 1987, 1991, and
2002, respectively.
He is currently an Associate Professor with the
Department of Mechanical Engineering, IST, and a
Senior Researcher with the Associated Laboratory of
Robotics and Systems in Engineering and Science.
He is the author or coauthor of more than 25 journal
papers and 100 conference communications. He has
participated in more than 20 European and Portuguese research projects over
the last 25 years. His research interests are in the area of autonomous robotic
vehicles with a focus on the elds of estimation, sensor fusion, navigation,
positioning, and industrial automation.
C. Guedes Soares received the M.S. and Ocean
Engineer degrees from the Massachusetts Institute of
Technology, Cambridge, in 1976, the Ph.D. degree
from the Norwegian Institute of Technology, Trond-
heim, Norway, in 1984, and the Doctor of Science
degree from the Technical University of Lisbon,
Lisbon, Portugal, in 1991.
He is a Professor of naval architecture and marine
engineering and the President of the Centre for Ma-
rine Technology and Engineering (CENTEC), which
is a research center of the Technical University of
Lisbon that is recognized and funded by the Portuguese Foundation for Science
and Technology. He has coauthored more than 400 journal papers and 600
conference papers and has been involved in more than 60 international research
projects and 20 national projects.