You are on page 1of 11

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/273123985

Assessment of ship collision estimation


methods using AIS data

Conference Paper · October 2014


DOI: 10.1201/b17494-27

CITATION READS

1 300

3 authors, including:

A. P. Teixeira Carlos Guedes Soares


Technical University of Lisbon University of Lisbon
101 PUBLICATIONS 997 CITATIONS 1,614 PUBLICATIONS 17,419 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

MSC research Project at Southampton university sponsored by Houlder Offshore View project

Smart control of ship propulsion systems View project

All content following this page was uploaded by Pedro Silveira on 18 August 2015.

The user has requested enhancement of the downloaded file.


Assessment of ship collision estimation methods using AIS data
P. Silveira, A. P. Teixeira & C. Guedes Soares
Centre for Marine Technology and Engineering (CENTEC), Instituto Superior Técnico,
Universidade de Lisboa, Portugal

ABSTRACT: The paper reviews different models that have been proposed for estimating ship collision can-
didates. The method based on the collision diameter proposed by Fujii and Pedersen is first used to assess the
number of collision candidates, using AIS data received from ships off the coast of Portugal during one
month. A new method is proposed to determine the number of collision candidates based on the available AIS
messages. In this method pairs of ships are modelled as rectangles, using length and breadth information
available in AIS messages. These rectangles are projected onto a line perpendicular to their relative velocity,
(using course, heading and speed information from AIS) and the pair is considered a collision candidate if
there is overlapping in the projections. A parametric study is performed to assess the contribution of im-
portant parameters of this method on the collision candidates estimates. The objective is to investigate the ca-
pability of the method for better predictions of collision candidates based directly on AIS data.

1 INTRODUCTION candidates, using one month of AIS data received


from ships off the coast of Portugal. The objective is
Over the years, concerns with maritime traffic regu- to identify the situations where collision avoidance
lation resulted in the introduction of Traffic Separa- actions prevented the collision. This reduces the col-
tion Schemes (TSS). The first TSS was introduced in lision candidates to values that would reflect the ac-
the Dover Strait in 1967 with the main objective of tual situations in which ships have taken collision
avoiding head on collisions. It was made mandatory avoidance actions in real collision scenarios.
by the International Maritime Organization (IMO) in
1971. Since then, many more TSSs have been intro-
duced in different parts of the world. These routeing 2 COLLISION ASSESSMENT MODELS
systems have been combined with traffic monitoring
by Vessel Traffic Services (VTS) in an attempt to An important concept when addressing ship colli-
improve the safety of navigation and reduce the sion problems is the concept of ship domain. Fujii &
number of ship collisions. The introduction of the Tanaka (1971) developed models to determine the
Automatic Identification System (AIS), intended to traffic capacity of a route or waterway which used
enhance the safety of navigation and situational the concept of ship domain, as an area around the
awareness and has enabled the collection of large vessel which the navigator would like to keep free of
amounts of data that may be used for studying mari- other vessels, for safety reasons. Goodwin (1975) al-
time traffic and associated risks, in particular ship- so used this concept in collision assessment models.
to-ship collision risk. Several shapes have been proposed for ship do-
Several models have been proposed to estimate the mains, such as circular, elliptical or polygonal.
frequency of ship collisions (e.g. Pedersen 2010; Fuzzy ship domains, such as proposed by Pie-
Montewka et al. 2010, Kaneko 2013). All models trzykowski et al. (2008) have introduced a higher
depend on a method to determine the occurrence of a level of sophistication to the concept, by creating
collision. Different methods may take into account different levels of safety within the domain, rather
length, breadth, speed, course difference, ship dy- than using a binary classification (safe or unsafe) for
namics, or combinations of these parameters. With a given situation. Ship domain models depend on the
the advent of AIS, large amounts of data have be- definition of the size of the domain, which may pre-
come available to develop such models (Silveira et sent significant variations (Wang et al. 2009).
al 2013). Most collision assessment models are based on the
This paper proposes a new method for estimating work of Fujii et al. (1971), and MacDuff (1974),
the ship collision candidates. A parametric study is which was later improved by Pedersen (2010).
carried out to analyse the influence of important pa- These authors calculate the frequency of collision as
rameters of the model on the assessment of collision
 =   (1) Kaneko (2013) considered that Pedersen’s (2010)
model was not satisfactory in cases where the cross-
where NG is the number of geometric collision can- ing angle is small, and proposed a method that cate-
didates during a time period (which is the number of gorizes the crossing angle by its sharpness into two
pairs of ships on a collision course during that time types. The method improves the estimation of the
period) and PC is the causation probability, defined number of collision candidates in crossings with
as the probability that a pair of ships which are on a small angles.
collision course (a collision candidate) fails to avoid
a collision.
Fujii et al. (1970) developed a model to determine
the number of geometric collision candidates from
ships sailing in crossing waterways. The model re-
lies on the calculation of a collision diameter, which
depends on the length, breadth and relative velocity
of ships using crossing waterways, and also on the
crossing angle of the waterways. The collision di-
ameter is calculated using the expression:
() () () ()
    ()
=  +  1 −  ∙


()  #/ ()  #/


 (#) 

! " + %1 − & ∙ 
' ( (2) Figure 1: Collision diameter according to Pedersen (2010)

where i and j are ship classes, Li(1) is the length of Montewka et al. (2010) developed a method that
class i ships on waterway 1, Lj(2) is the length of replaces the geometric collision diameter by a Min-
class j ships on waterway 2, Bi(1) is the breadth of imum Distance To Collision (MDTC), a critical dis-
class i ships on waterway 1, Bj(2) is the breadth of tance under which collision avoidance actions can-
class j ships on waterway 2, Vi(1) is the average ve- not prevent the collision from occurring. The value
locity of class i ships on waterway 1, Vj(2) is the av- of this minimum distance depends on the crossing
erage velocity of class j ships on waterway 2, Vij is angle and on the ship type, and is calculated based
the relative velocity between class i and class j ves- on a ship dynamics model, improving traditional
sels and θ is the crossing angle of the two water- methods that rely on a constant critical distance to
ways. Vij is computed using the expression all vessels and any crossing angle. The method was
applied for 3 types of vessels (container carrier, pas-
 
) = *+)
(#) () (#) ()
, + +) , − 2) ) ./ (3) senger ship and tanker), 9 meeting scenarios, 17
crossing angles (from 10º to 170º) and 4 types of
manoeuvres conducted by both vessels to avoid col-
Pedersen (2010) developed Fujii’s model to in- lision. The method relies on the assumption that the
clude the probabilistic distribution of ships along the vessels are under way at full sea speed and both start
width of the waterways. The number of collision evasive actions simultaneously.
candidates is taken as: Goerlandt & Kujala (2011) developed a method
3 .3
0 = ∑ ∑ 2Ω 278 ,8 (6 ) ∙
to determine the number of collision candidates that
(#)
5
: 
() () relies on a simulation of ship traffic. The number of
∙
collision candidates is computed using an algorithm
()
5 76 :) <=∆? (4) that determines, for each combination of traffic
events (a traffic event is defined as one voyage of
where Q1i is the number of class i ships using wa- one ship from a given point of departure to a specific
terway 1 during time period ∆t, Q2j is the number of destination along a predefined route), in how many
class j ships using waterway 2 during time period moments in time the contours of simulated ships
∆t, zi is the distance of class i ships to the centre of overlap.
waterway 1, zj is the distance of class j ships to the All the methods described are based on the ships
centre of waterway 2 and f is the lateral distribution position, course and speed, which will lead to colli-
of traffic using a waterway. sion in case no changes are made. There are a num-
Figure 1 shows the collision diameter used by Fu- ber of collision avoidance systems that use similar
jii et al. (1970) and Pedersen (2010) to determine concepts to identify the situations in which it is nec-
collision candidates. essary to change course and speed to avoid collision,
including in close quarters, as for example described
in Perera et al. (2011, 2012), as well as in references across waterways were modelled from AIS data for
therein. However, despite the existence of that type all vessel types and meeting scenarios considered.
of devices and on the human actions that will also be This was achieved by defining three counting gates:
present in many cases, the probability of collision one gate counted the number of vessels arriving and
continues being proportional to the number of ships departing from the Gulf of Finland, another gate
in this risk situation, in which case the present anal- counted N-S traffic and the remaining gate counted
ysis is justified. E-W traffic.
Mou et al. (2010) performed a statistical analysis
of ships involved in collisions, identifying the corre-
3 USE OF AIS DATA IN SHIP COLLISION lation of CPA (Closest Point of Approach) with
MODELS ship’s size, speed and course, using AIS data from a
Traffic Separation Scheme (TSS) off the port of Rot-
From 31 December 2004 all vessels over 300 GT terdam. A collision risk assessment (with the term
engaged in international voyages, all cargo vessels “risk” standing for probability of collision in the
over 500 GT and all passenger vessels are required context) was performed using SAMSON (Safety As-
to be equipped with AIS. This system allows auto- sessment Models for Shipping and Offshore in the
matic exchange of information between stations North Sea), calculating the average basic collision
(ships and coastal), using VHF radio waves. There risk in the area based on ships’ type, size, age, flag,
are 27 message types defined in ITU (International weather conditions and location. This basic collision
Telecommunication Union) recommendation risk was then multiplied by factors related to TCPA,
M.1371-5, and two classes of shipboard equipment: CPA and encounter angle, resulting in a dynamic
class A (used mainly by commercial vessels) and risk of collision value intended to be used in real
class B (used mainly by fishing vessels and pleasure time decision support systems.
craft). The reporting intervals of class A equipment Goerlandt & Kujala (2011) used information that
vary between 2 seconds and 10 seconds (depending resulted from a detailed AIS data analysis as input
on the ship’s speed and rate of turn) if the ship is not for their simulation. Individual vessel trajectories
moored or at anchor. If the ship status is moored or were grouped into routes corresponding to voyages
at anchor, the reporting interval is 3 minutes, unless between two ports. From these routes a number of
the speed is greater than 3 knots, which sets the re- distributions were extracted, such as departure time
porting interval to 10 seconds. The message types (which in this case cannot be satisfactorily described
used in this study were position reports (message by a stationary Poisson model), route, ship type, ship
type 1, 2 and 3) and static and voyage related data dimensions and ship speed.
(message type 5) transmitted by class A equipments. Silveira et al. (2012) characterized the maritime
Information contained in position reports includes traffic and performed a statistical analysis of the
date/time, Maritime Mobile Service Identity traffic using a TSS off the coast of Portugal based on
(MMSI) number, navigation state, rate of turn, speed AIS data. Moreover an algorithm was developed to
over ground, position accuracy, latitude, longitude, assess the risk profile and the relative importance of
course over ground and heading. Static and voyage routes associated with ports.
related data messages include information on MMSI Using the same AIS data set, Silveira et al. (2013)
number, IMO number, ship’s name, destination, estimated the number of collision candidates based
ETA, callsign, type of ship, length, breadth and on predictions of future positions of ships and on
draught. the Fujii’s et al. (1970) and Pedersen’s (2010) defi-
AIS data are becoming an increasingly important nition of collision diameter. Values of causation
source of information for maritime traffic studies. probability suggested in several studies were used to
Aarsæther & Moan (2009) applied computer vision calculate the expected number of collisions which
techniques to automatically separate AIS data in or- were then compared with the number of collisions
der to obtain traffic statistics and prevailing features that have occurred between 1997–2006, registered
of traffic, enabling the production of a simplified and maintained by the Portuguese Maritime Authori-
ship traffic model. The computer vision techniques ty.
are used assuming that there are well defined ma- The reliability of information provided by AIS
noeuvre patterns, which can be identified by analyz- ship stations has also been object of numerous stud-
ing traces of the ship positions. The method is found ies. Harati-Mokhtari (2007) exposes various errors
to yield good results in normal navigation scenarios, detected in AIS messages, that can be originated by
while the accuracy in an inner harbor area is limited wrong static information inputted during installation
due to a very complex manoeuvring behaviour. and/or by wrong voyage information inputted while
Montewka et al. (2010) used AIS information to operating the equipment.
model the vessel traffic profiles used to develop
their MDTC method. Vessels’ course, speed, length
and breadth, and the lateral distribution of traffic
4 ASSESSMENT OF COLLISION
CANDIDATES

The assessment of the number of collision candi-


dates is carried out based on one month AIS data
collected from ships in an area around the TSS off
Cape Roca (Figure 2) using two different methods:
1. Collision Diameter Method: compares the Clos-
est Point of Approach (CPA) with the Collision
Diameter (CD), as defined by Fujii et al. (1970)
and Pedersen (2010).
2. Projection Method: is based on the overlapping
of the projections of ships modelled as rectangles
(using length and breadth information from AIS
messeges) onto a line perpendicular to their rela-
tive velocity vector.

Figure 3: AIS message types 1, 2 and 3 in the studied area, be-


tween 09/07/2008 and 09/08/2008.

L
2
B

C D B
2
Figure 2: TSS off Cape Roca and Berlengas' area to be avoided Figure 4: Reference point for AIS position and overall dimen-
(2008). Source: Portuguese Hydrographic Office (IH) sions of ship. Note that dimension B on the left is not related to
B/2 (half-breadth) on the right.

Figure 3 shows a plot of AIS traffic in the studied CPA is the smallest distance at which the correct-
area during the time period between 09/07/2008 and ed positions if two ships can be from each other,
09/08/2008. computed based on their positions at a given mo-
For both methods it is necessary to correct the po- ment and assuming constant courses and speeds. To
sition contained in AIS messages, which is the posi- apply method 1, CPAs for all possible pairs of ships
tion of the GPS antenna, to a position located at half- were computed for every second of data recorded
length and half-breadth. This can be achieved know- and compared with collision diameter computed for
ing the ship’s heading and dimensions A, B, C and the same pairs of ships. If CPA was less than the
D shown on Figure 4, which are transmitted in AIS collision diameter and TCPA (Time to Closest Point
type 5 messages. of Approach) was positive (the CPA will occur in
the future), the pair was considered a collision can-
didate.
Situations may occur for which the CPA is small- BC EF = @ABC ⋅ )KBCL (7)
er than the collision diameter but the ships do not
with @ABC as the position vector of the center of rec-
collide. The heading information that can be trans-
mitted in AIS position reports may be used to en-
hance the classification of collision candidates. tangle B relative to the center of rectangle A. The
Method 2 is proposed with this intent. overlapping occurs when:
CPQR BPQR
To apply method 2 it is necessary to determine
the direction perpendicular to the relative velocity of BC EF ≤

(8)
two vessels. The relative velocity vector is computed
using COG (Course Over Ground) and SOG (Speed For both methods, if the distance between the two
Over Ground) information from both vessels, con- ships is greater than 20 miles the pair is not consid-
tained in AIS messages. The rectangles centred on ered a collision candidate because it is assumed that
the corrected positions, with the principal dimension no collision avoidance action will be taken when the
aligned with the ship’s heading, with length equal to ships are that far from each other. If the pair is in-
the ship’s length and width equal to the ship’s deed a collision candidate is classified as such when
breadth are then projected onto a line perpendicular the distance is less than 20 nautical miles (nm).
to the relative velocity vector, as shown in Figure 5
and Figure 6. If the projections overlap, the pair is
considered a collision candidate (Figure 6). This B
procedure was repeated for every second of data (or Hb
each time the date/time field of one AIS message CB
was different from the one on the previous message).
C A V BA

B CA

HB HA

CB

V BA
C A
A

CA
HA

A
Overlap of projections

Figure 6: Projection of rectangles - collision candidate

The two methods were applied using AIS mes-


sages transmitted between 09/07/2008 and
Figure 5: Projection of rectangles - no collision candidate 09/08/2008, by ships in an area bounded by meridi-
ans 009º 28’ W and 010º 30’ W and by parallels 38º
The relative velocity vector is: 16’ N and 39º 50’ N. All collision candidates for

@ABC = DAB − DAC


which the collision would occur outside of the des-
) (5) ignated area were disregarded.
Pairs of ships were classified several times as col-
The length of rectangle A’s projection is given lision candidates until one or both of them altered
by: course or speed. The algorithms for both methods
classify the pair as multiple collision candidates for
HC|| ⋅ )KBCL MC + G
=EF = G HCL ⋅ )KBCL C (6) the same possible collision event, so it is necessary
to delete repeated collision candidates. The algo-
with GHC|| as a unit vector parallel to ship A’s head- rithm used to delete the repeated candidates com-
ing, GHCL as a unit vector perpendicular to ship A’s pares MMSIs, COGs, SOGs of both ships and the
heading, )KBCL as a unit vector perpendicular to the time at which the collision candidate occurred with
relative velocity vector, MC as the length of ship A past collision candidates to determine if the collision
and C as the breadth of ship A. The projected dis- candidate event is already recorded. If so, the most
tance between the centers of rectangles A and B is: recent collision candidate is kept and the others are
deleted.
The projection method was then applied introduc- ment they are classified as collision candidates until
ing uncertainty in the position included in AIS mes- the moment they stop being collision candidates). It
sages. If the position of the centre of the rectangle is was found that for 27.3% of the candidates this time
within SET meters of the position included in the interval was less than one minute. In fact, one mi-
AIS message, the shape of all possible positions of nute is not enough time to analyze and perceive a
situation as a potential collision and therefore the ac-
the rectangle is depicted in Figure 7 a), with L and B
tion taken would not be classified as a collision
as the length and breadth of the ship, respectively.
with ∆?V of less than one minute were disregarded.
avoidance action. As a consequence, all candidates
However, to simplify the application of the meth-
od, the shape of all possible positions of the rectan- Table 2 shows the number of collision candidates af-
gle was assumed to be as seen in Figure 7 b), in- ter this correction.
creasing both the length and the width of the
rectangle by 2SET . Projection SET (U)
Method 0 10 20 30
B+2 e pos Collision
3429 3758 3946 4081
candidates
B
Collision diameter CD / 2 CD
Method
Collision
4188 4613
candidates
Table 2: Collision candidates, corrected by removing those
L L+2 e pos with ∆WX under 1 minute

As stated before, if the initial distance between


the two ships is greater than dmax=20 nautical miles
the pair was not considered a collision candidate be-
cause it is assumed that no collision avoidance ac-
a) b) tion was taken when the ships are that far from each
other. However, this limit may also lead to the clas-
Figure 7: Position error sification of “false positives”, pairs of ships that are
considered a collision candidate and subsequently
The collision diameter method was also applied change their course not as a collision avoidance ac-
for half of the collision diameter (CD/2), shown in tion but as a planned course alteration according to
Figure 1 and for the full collision diameter. The re- their route plan. Reducing the value of this maxi-
sults are shown in Table 1. mum distance (dmax) is expected to reduce the num-
ber of “false positives”, but may also prevent colli-
Projection SET (U) sion candidates that take evasive actions very early
method 0 10 20 30 from being identified.
Collision A sensitivity analysis was performed on the clas-
4719 5033 5222 5347
candidates sification algorithm, using both methods, to deter-
mine how the described maximum distance parame-
Collision diameter CD/2 CD ter (dmax) affects the number of collision candidates.
method The results are shown on Table 3.
Collision
5340 5656
candidates
Table 1: Collision candidates estimated using the projection dmax (Maximum distance in nm)
and collision diameter methods 5 10 15 20
Collision
“False” collision candidates may result from situ- diameter 1962 3237 4114 4613
ations in which a ship is altering its course and is in method
a collision course with another ship but only during Projection
brief moments. To eliminate some of these “false” method
1387 2390 3052 3429
collision candidates (ships that were on a collision
course and change their courses or speeds but not as Table 3: Number of collision candidates for different maximum
a collision avoidance action) the collision candidates distances (dmax)
were sorted by the time interval during which they
were collision candidates (∆?V ) (i.e. from the mo-
The histograms in Figure 8 and Figure 9 were 20 minutes, using AIS messages from a different
produced from information associated with the colli- (much larger) geographical area. Although it is not
sion candidates. possible to directly compare the results presented
Figure 8 shows the distribution of collision can- here with the results published by Silveira et al.
didates according to the value of TCPA when they (2013), the method used in that work is equivalent to
stopped being collision candidates, that is, when the collision diameter method presented here, with
some action was taken that resulted in that pair of the difference that the collision candidates classifica-
ships not being in a collision course anymore. The tion was made for every second of data in this study,
candidates were classified with the projection meth- whereas for the previous study (Silveira et al.
od for a maximum distance of 20 nautical miles. (2013)) the classification was made for every 30
seconds of data, due to limitations in computational
power given the much larger number of AIS mes-
sages. In Figure 8 the number of candidates peaks
are between TCPAs of 20 and 40 minutes, while in
Figure 9 the number of candidates peaks are be-
tween distances of 1 and 4 nautical miles. This is
consistent with the expected behaviour of ship mas-
ters when performing collision avoidance actions.
However, there are differences between head-on,
crossing and overtaking situations. Head-on situa-
tions are defined as being those with a course differ-
ence of 170º or higher; overtaking situations are de-
fined as being those with a course difference not
exceeding 10º; and crossing situations are all the re-
maining (Montewka et al. 2010). This classification
Figure 8: Distribution of collision candidates by final TCPA,
projection method, maximum initial distance of 20 nm
of collision situations is not necessarily coincident
with the definitions of the COLREGs (International
Regulations for Preventing Collisions at Sea). Par-
ticularly, in an overtaking situation, if the relative
speed is very small, final distances tend to be lower
and final TCPAs higher than in a crossing situation.
This can be observed by comparing Figure 10 with
Figure 11 and Figure 12 with Figure 13. The number
of candidates classified as head-on, crossing and
overtaking situations is shown on Table 4. The ex-
pected values of final distance and final TCPA are
shown on Table 5.

Projection Method
Head-on Crossing Overtaking
Collision
32 1082 2315
Figure 9: Distribution of collision candidates by final distance, candidates
projection method, maximum initial distance of 20 nm
Collision diameter method
Head-on Crossing Overtaking
Figure 9 shows the distribution of collision can- Collision
didates according to the final distance between the candidates
73 1343 3197
two ships when they stopped being a collision can-
didate. It is clear that the number of collision candi- Table 4: Collision candidates by type of collision situation,
dates increases as the distance decreases, indicating maximum distance of 20 nm
some relation with collision avoidance actions.
However, it is not possible to define an exact limit of Head-on Crossing Overtaking Total
distance under which there are no false collision Final distance
candidates identified. 8.65 8.47 5.88 6.72
(nm)
Silveira et al. (2013) estimated the number of col- Final TCPA
lision candidates using the collision diameter meth- 20.0 47.4 69.5 62.6
(minutes)
od with a TCPA limit of 20 minutes, excluding all
candidates that took collision avoidance actions Table 5: Expected values of final distance and final TCPA,
projection method, maximum initial distance of 20 nm
when the remaining time before a collision was over
Figure 10: Distribution of collision candidates by final TCPA, Figure 13: Distribution of collision candidates by final dis-
projection method, maximum initial distance of 20 nm, cross- tance, projection method, maximum initial distance of 20 nm,
ing situations overtaking situations

The values in Table 5 are useful to compare final


distances and final TCPAs in different collision situ-
ations, however they do not reflect the normal be-
haviour of ships performing collision avoidance ac-
tions (experience determines that the majority of
ships involved in a crossing collision avoidance situ-
ation do not take action when they are 8 or 9 miles
from the other ship, or when there are still 47
minutes before the collision). The reason for this de-
viation from reality is the number of “false” colli-
sion candidates, ships that were on a collision course
and stopped being not because of a collision avoid-
ance action but because of scheduled course altera-
tion or some other reason.
From the results of the collision candidates ob-
Figure 11: Distribution of collision candidates by final TCPA, tained using the projection method, it is possible to
projection method, maximum initial distance of 20 nm, over- determine that for 9 pairs of ships the minimum dis-
taking situations tance was less than the collision diameter. Using the
collision diameter candidate classification method,
10 pairs of ships were in a situation where the final
distance between their reference points (the distance
when they stopped being collision candidates) was
less than the collision diameter. In these close quar-
ters situations collision was imminent.
Figure 14 shows the geographical distribution of
collision candidates classified with the projection
method (in red) and the 9 referred near collision sit-
uations identified with the same method (in blue).

Figure 12: Distribution of collision candidates by final dis-


tance, projection method, maximum initial distance of 20 nm,
crossing situations
The heading information depends on the correct
installation of AIS equipment and on the connection
to an external device, typically a gyro compass or a
magnetic compass. If there is no heading infor-
mation, the special code ‘511’ is transmitted in the
heading field. About 4.9% of used messages of
types 1, 2 and 3 had no heading information. In
these cases the COG was taken as a good approxi-
mation of the heading, and used to compute the pro-
jection. The information on ship’s dimensions de-
pends on the correct installation of AIS equipment,
given that this is static data and is not altered by the
crew. About 6.3% of the used type 5 messages had
no information about ship’s dimensions.
In future work, the quality of the ship dimensions
information included in AIS messages may be eval-
uated by comparing it with information in ship data-
bases, such as the national maritime database main-
Figure 14: Collision candidates and near collisions, projection tained by DGRM (Direcção Geral de Recursos
method, maximum initial distance of 20 nautical miles
Naturais, Segurança e Serviços Marítimos) or other
public or private databases. The quality of the head-
ing information may be evaluated by comparing it
5 POSITION ACCURACY AND ERRORS IN AIS with course and speed information, and also by
INFORMATION
comparing those values with heading, course and
speed of nearby vessels. This information may be
The quality of the results obtained by the applica- used to compute average set and drift and estimate
tion of the projection method, to determine the num- the expected difference between heading and course
ber of collision candidates, depends on the precision for a given ship, excluding the headings that fall out-
of the position and heading information and on the side a reasonable range.
reliability of the ship dimensions included in AIS
messages. AIS messages contain some information
about the origin and accuracy of the position being 6 CONCLUSIONS
transmitted. Message types 1, 2 and 3 contain a posi-
tion accuracy flag, which indicates if the position is The proposed method of projecting rectangles (as
accurate within 10 meters or not. Message type 5 representations of ships) to determine the number of
contains a field which indicates the type of electron- collision candidates from AIS records has the poten-
ic position fixing device used to obtain the position. tial to yield more precise results than the method us-
The value of position accuracy found on the AIS ing the collision diameter, as shown by the number
messages used is shown on Table 6. About 28% of of collision candidates computed, even when a posi-
the positions included in position reports are accu- tion error was introduced. Both methods depend on
rate within 10 meters. Assuming that all these posi- the quality of AIS data. The projection method re-
tions have an error of 10 meters and that the posi- quires knowledge of the heading of ships in addition
tions included in all other messages have an error of to all information required to apply the collision di-
30 meters, using the data presented in Table 2, the ameter method, although when the heading infor-
number of collision candidates can be estimated as mation is not available the ship’s course may be
3990 (projection method). These assumptions are
used as an approximation.
made without rigorous information of position accu-
The projection method is also suitable for use in
racy in the area, and only to demonstrate that a GPS
real-time collision avoidance systems (always keep-
position accuracy correction factor may be intro-
ing in mind the limitations of AIS) and in traffic
duced to the estimation of the number of collision
simulations. The collision candidates obtained by
candidates.
both methods are influenced by “false” collision
candidates, i.e. pairs of ships that stopped being col-
Position accuracy flag
lision candidates because of a scheduled or other
Accuracy Number of AIS messages
type of course or speed alteration not related to colli-
≤ 10 m 3831268 sion avoidance. More realistic estimates of collision
> 10 m 9698232
candidates (i.e. with less “false” candidates) can be
Table 6: Position accuracy of AIS messages obtained by decreasing the maximum distance pa-
rameter (i.e. ignoring all candidates which are at a
greater distance than this value), by ignoring candi-
dates with a TCPA under a certain limit, or a combi- ing, State Estimation, and Trajectory Prediction. IEEE
nation of both. Transactions on Intelligent Transportation Systems,.
13(3):1188-1200.
Pietrzykowski, Z. (2008). Ship’s fuzzy domain – a criterion for
navigational safety in narrow fairways. Journal of Naviga-
ACKNOWLEDGMENTS tion, 51, 499-514.
Silveira, P.A.M., Teixeira, A.P. & Guedes Soares, C. (2012).
The authors would like to thank DGRM (Direcção Analysis of maritime traffic off the coast of Portugal, Mari-
Geral de Recursos Naturais, Segurança e Serviços time Engineering and Technology. Guedes Soares, C. Gar-
batov Y. Sutulo S. & Santos T. A., (Eds.). Taylor & Francis
Marítimos) for providing the AIS data used in this Group, London. pp. 35-41.
study. The contribution of the second author has Silveira, P.A.M., Teixeira, A.P. & Guedes Soares, C. (2013).
been provided under the scope of the project “IRIS- Use of AIS Data to Characterise Marine Traffic Patterns
Project Risk Management: Improving Risk Matrices and Ship Collision Risk off the Coast of Portugal, Journal
Using Multiple Criteria Decision Analysis”, funded of Navigation, Vol. 66, pp. 879–898
Wang, N., Meng, X., Xu, Q. & Wang, Z. (2009). A unified an-
by the Portuguese Foundation for Science and Tech- alytical framework for ship domains. Journal of Naviga-
nology (Fundação para a Ciência e a Tecnologia - tion, 62, 643-655.
FCT), contract PTDC/EGE -GES/119230/2010.

REFERENCES

Aarsæther, K. & Moan, T. (2009). Estimating navigation pat-


terns from AIS. Journal of Navigation, 62 (4): 587-607.
Fujii Y., Tanaka K. 1971. Traffic capacity. Journal of Naviga-
tion, 24, 543-552.
Fujii, Y. & Shiobara, R. (1971). The Analysis of Traffic Acci-
dents. Journal of Navigation 24 (4): 534-543.
Fujii, Y., Yamanouchi, H. & Mizuki, N. (1970). On the funda-
mentals of marine traffic control. Part 1: Probabilities of
collision and evasive actions. Electronic Navigation Re-
search Institute Paper 2: 1-16.
Goerlandt, F. & Kujala, P. (2011). Traffic simulation based
ship collision probability modeling. Reliability Engineering
and System Safety, 96: 91-107.
Goerlandt, F., Montewka J., Lammi, H., & Kujala, P. (2012),
Analysis of near collisions in the Gulf of Finland. Advances
in Safety, Reliability and Risk Management, Bérenguer,
Grall & Guedes Soares (Eds.). Taylor & Francis Group,
London. pp. 2880-2886.
Goodwin, E.M. (1975). A statistical study of ship domains.
Journal of Navigation, 57, 329-341.
Harati-Mokhtari, A. (2007). Automatic Identification System
(AIS): Data Reliability and Human Error Implications.
Journal of Navigation, 60: 373-389.
Kaneko, F. (2013). An improvement on a method for estimat-
ing number of collision candidates between ships, Collision
and Grounding of Ships and Offshore Structures, –
Amdahl, Ehlers & Leira (Eds), Taylor & Francis Group,
London, pp. 27-37.
Macduff, T. (1974). The probability of vessel collisions. Ocean
Industry 9(9): 144–148.
Montewka, J., Hinz, T., Kujala, P. & Matusiak, J. (2010),
Probability modeling of vessel collisions, Reliability Engi-
neering and System Safety 95, pp. 573-589.
Mou, J.M., Tak, C. & Ligteringen, H. (2010), Study on colli-
sion avoidance in busy waterways by using AIS data,
Ocean Engineering 37, pp. 483-490.
Pedersen, P.T. (2010), Review and application of ship collision
and grounding analysis procedures, Marine Structures 23,
pp. 241-262.
Perera, L.P., Carvalho, J.P. and Guedes Soares, C. (2011),
Fuzzy-logic based decision making system for collision
avoidance of ocean navigation under critical collision con-
ditions. Journal of Marine Science and Technology.
16(1):84–99.
Perera, L.P., Oliveira, P. & Guedes Soares, C. (2012). Mari-
time Traffic Monitoring based on Vessel Detection, Track-

View publication stats

You might also like