International Journal

on Marine Navigation

and Safety of Sea Transportation

Volume 4

Number 3

September 2010

303

1 INTRODUCTION

Analysis of ship traffic receives focus as the aware-

ness of the risk it poses to the environment is in-

creased. The analysis is not only motivated by the

desire to quantify risk, but also to understand the ef-

fect of changes to the fairway and to propose im-

provements to harbor areas and inland waterways.

Analysis of ship traffic has been hindered by a scar-

city of data, requiring specialized installations or

equipment for data collection. This scarcity has

prompted studies that rely on synthetic ship maneu-

vering data from simulators (Hutchinson 2003) and

(Merrick 2003). While simulator studies provides

valuable insights, through high sample rates, con-

trolled environment and absence of noise, they make

compromises on either the number of passages with

the use of human operators in full-mission simula-

tors, or on accuracy by eliminating the human ele-

ment and relying on fast-time simulators with auto-

pilot algorithms.

We will in this paper show how the introduction

of the Automatic Identification System (AIS) for

ships can help in both providing a readily available

data source for traffic analysis, and how analysis of

this data can be employed to generate statistics of

traffic conditions, estimate maneuver plans and pa-

rameters as inputs to fast-time simulator studies

The use of AIS in marine traffic analysis is not a

new concept. (Gucma 2007) used AIS data to esti-

mate the occurrence of accidents in the Baltic Sea by

identifying the major traffic flows using AIS records

of journeys. Little work has been done to apply AIS

to analysis on the scale of maneuvers in a smaller or

constrained area to derive the exhibited maneuver

patterns. The area around the harbor of Risavika in

southwestern Norway was selected as a case study.

This area was selected since the presence of island

formations and the coastline should impose a struc-

ture on the ship traffic. Information about all the

navigational markings in the area was obtained from

the Norwegian Hydrographic Service and contained

data for position, type and identifiers for all publicly

maintained navigational aids in the area. The area

with AIS position reports and navigational markings

indicated is shown in Figure 1.

Figure 1: Risavika harbor with AIS position reports and navi-

gation markers indicated

Computer Vision and Ship Traffic Analysis:

Inferring Maneuver Patterns From the

Automatic Identification System

K. G. Aarsæther & T. Moan

Norwegian University of Science and Technology, Trondheim, Norway

ABSTRACT: The Automatic Identification System has proven itself as a valuable source for ship traffic in-

formation. Its introduction has reversed the previous situation with scarcity of precise data from ship traffic

and has instead posed the reverse challenge of coping with an overabundance of data. The number of time se-

ries available for ship manoeuvring analysis has increased from tens, or hundreds, to several thousands. Sift-

ing through this data manually, either to find the salient features of traffic, or to provide statistical distribu-

tions of decision variables is an extremely time consuming procedure. In this paper we present the results of

applying computer vision techniques to this problem and show how it is possible to automatically separate

AIS data in order to obtain traffic statistics and prevailing features down to the scale of individual manoeu-

vres and how this procedure enables the production of a simplified model of ship traffic.

304

The small-scale analysis benefits from the large

volume of data available, but the size of the data sets

involved makes analysis more demanding. At this

level the analysis method must take into account the

alternate routes through the area and the possibility

of harbors.

The introduction of AIS has replaced the previous

scarcity of ship traffic and maneuver data with an

overabundance. Whereas one previously had to con-

struct limited shore based measurement systems

with limited lifespan or rely on data from a selected

set of vessels with logging equipment, AIS provides

a continuous stream of information of the position

and speed all AIS enabled vessels in range. The sys-

tem provides position and speed updates on prede-

fined intervals depending on vessel speed and ma-

neuver situation with a sample-rate varying from 3

seconds for high speed or turning vessels to 15 min

for ships at anchor. The instantaneous information

density varies naturally with the traffic density of the

area, but if one considers past data, the amount of in-

formation to be shifted through and analyzed is con-

siderable. The ability to analyze, and the capabilities

of the techniques employed, will determine the qual-

ity of information about ship traffic extracted from

this new source of historic ship traffic data.

On this background we present a method based

on computer vision techniques, which is capable of

handling this increase in available data.

2 METHOD

Analyzing ship traffic is a two-stage process where

the first task is to transform the collected data into a

form that eases the final analysis. A method for

transformation of AIS data frames to a collection of

maneuvers presented is comprised of following

stages:

− Reconstruction of vessel specific time-series from

AIS data

− Sorting of time-series from geometric similarity

of the position trace

− Subdivision of the geometric similar groups into

groups with the same direction of travel

This process produces groups of time-series with

similar maneuver patterns and direction of travel

well suited for generation of statistics and further

analysis. Further analysis of these groups can in-

clude:

− Traffic properties such as distribution of vessel

velocity and spread

− Estimation of maneuver sequence and parameter

statistics

− Estimation of the most probable navigation aid

used for maneuver transitions

2.1 Model for ship maneuvers

The ship maneuvering process is represented as a

sequence of basic maneuvers. The basic maneuvers

are instantiated and appear as a recognizable maneu-

ver pattern. The most basic subdivision of maneuver

patterns is the distinction between constant course

and course changing maneuvers. While these catego-

ries can contain variations in the strategies employed

to obtain the desired result, the two groups are repre-

sents the simplest geometric model is the model ship

maneuvering.

2.2 AIS data collection

Data frames from the Norwegian AIS stations in the

area around the harbor of Risavika were collected

for three months from April to June 2006. AIS data

frames are marked with a time stamp and the vessel

specific MMSI number and contains the vessels in-

stantaneous position and speed if available.

The data was ordered by MMSI number and time

to recreate the time-series for each vessel. The time-

series was then split at significant discontinuities in

time to handle the cases of vessels leaving the stud-

ied area or coming to rest in a harbor. The number of

AIS position reports in the area was 512,533 and the

position reports were reduced to 2763 time-series

2.3 Grouping of time-series

Application of image registration techniques solves

the laborious task of grouping the time-series form

the geometric similarity of the position trace.

Image registration techniques (Zitova 2003,

Brown 1992) is applied in medical imaging and pro-

duction control, and can be explained as the process

of comparing images mathematically to produce an

objective measure of their similarity and to detect

the presence of a-priori known objects. These tech-

niques are well suited for sorting vessel trajectories

from their geometry as the position trace in isolation

forms a line in an otherwise empty space. The trace

of a vessels position can be transformed into the

form of a digital image by discretization of the re-

ported position. The studied area was dividend into

75x75m bins and the number of position in each bin

was counted and stored in a matrix for each time se-

ries. This is the representation used for grayscale

images in image analysis.

Application of image registration must account

for the possible differences in image resolution, rota-

tion and translation of the captured scene. These pa-

rameters are controlled due to the transformation of

remotely sensed data into an image with controlled

orientation and resolution, but the location of the

imprint of the individual vessel traces introduces an

unknown possible translation. An application of im-

305

age registration to group geometric similar tracks

must account for this displacement within each

group, but a global compensation will introduce er-

rors, as it will detect similar position traces of simi-

lar form, but of very different location.

To reduce processing time the sorting of position

traces was divided into a coarse and detailed analy-

sis. The coarse analysis simply looked at the corre-

spondence of track images without accounting for

possible translations. If two images were deemed

similar, future comparison was done with the mean

track of the two. The coarse analysis left a large

number of small groups, which were used as inputs

to the detailed analysis. The detailed analysis made

use of numerical optimization to find the optimum

level of similarity between the groups. The cross-

correlation between two images, where one has a

translation (u,v) in (x,y) direction was used as an ob-

jective function and is seen in Equation (1).

)),(,),(max(

),(),(

),(

∑ ∑

−−

∑ ∑

∑

−−

∑

⋅

x y

vyuxI

x y

yxT

vyu

xIyxT

vuCCR

(1)

where T=reference image matrix and I=the test

image matrix with translation (u,v). The cross-

correlation defined in Equation (1) is only valid for

integer values (u,v) so a 3D interpolation method

from (Vetterling 2007) was implemented to provide

a continuous formulation of the cross-correlation.

The interpolation routine allows standard numerical

optimization strategies, such as steepest descend, to

be applied to find the maximum correlation between

two images. The cross-correlation was used to refine

the grouping obtained by the coarse method by an it-

erative process where groups which showed a max-

imum correlation where combined.

The final operation on the sorted time series was

to split each geometrically similar group into direc-

tion specific groups by considering the angle be-

tween the start and end points of each time-series.

2.4 Group maneuver identification

The time-series reconstructed from AIS have heter-

ogeneous sample rates, within the geometric similar

group, and even within the individual time-series.

This necessitates a transfer of individual time series

data to a common representation, which compen-

sates for the variations in sample-rates. The proper-

ties of the time series group was estimated from 100

evenly spaced control points. The control points

were computed as the mean points of 100 evenly

spaced points for each time-series belonging to the

group. Mean perpendicular vectors to the mean path

were calculated in conjunction with the control

points and used to establish mapping of the time-

series indices to the control points by finding the in-

tersection between the time-series trace and the per-

pendicular vectors.

The sequence of maneuvers in a group was

tracked by the curvature of the vessels trajectory.

The curvature of the vessel trajectories in each time

series was computed by considering the x and y co-

ordinate as signals in the time (Aarsæther 2007) as

seen in Equation 2.

( )

2/3

xyyx





−

(2)

A polynomial was fitted locally to the x and y

signals in time to provide well-defined derivatives

for curvature calculation. The curvature of each time

series was transferred to the group by the index to

control point mapping. The group curvature is then

calculated from the median of the group curvature at

each control point. The individual turn and straight

sections of the group are identified by an ad-hoc

two-stage filtering based on statistics. The mean

value,

, and standard deviation,

, are calculated

and the points of the group curvature curve that falls

outside the region defined by

± 2

are defined as

belonging to a turn section,

and

are recalculated

for the remaining points and the process repeated

once more. Contiguous regions are identified as turn

and straight segments.

The indentified turns in the group curvature only

provide information about the median straight and

turn behavior, to extend the analysis to the parame-

ters of the maneuver model and to provide statistics

demands data for the identified sections from each

time-series. The translation between the turn sec-

tions of the median path to the individual time series

is not well defined as the map of positions to control

points. Variations in curvature can occur at different

positions along the path and it is the sequence of

maneuvers that is of interest. The turn sections of the

median curvature were isolated and transferred to an

image representation using the same procedure as

for the position trace. The image representations of

the individual turn sections were then matched to a

section of the time-series by optimization of the sim-

ilarity between the turn image from the group curva-

ture and the time-series curvature. This identifies the

locations of the turn sections in the individual time-

series and enables the extraction of statistics based

on the maneuver progression of the individual ves-

sels instead of relying on geometric areas or indices

from the group curvature to extract data.

2.5 Statistics of time-series groups

Statistics for each time-series group was calculated

at the intersection between straight ant turn sections.

The sections of the individual time-series sections

were transferred to the group sections by mapping

the turns in a time-series to the corresponding turn

numbers in the group. Variables were according to

section type

306

− Turn section: extreme, median & mean curvature

and median speed over section

− Straight section: average course angle over sec-

tion, offset from median path at both endpoints

and median speed over section

2.6 Identification of navigational aids

The identification of the most used navigational aids

is dependent on the location of the border points be-

tween the straight and circular sections of the ships

path in relation to the navigational markings in the

environment.

The identification of the most probable naviga-

tional aid is more error prone than processing of AIS

data since the result is directly influenced by the

choice of criterion used to identify the aid used in

each time-series. The identification criterion used is

based on ship-handling theory, where navigation

references are preferred if the bearing from the turn

initiation point to the reference is close to parallel

with the future course. The angle to all the naviga-

tional markings in the area was calculated for each

turn initiation, and the marking with a bearing clos-

est to the course at the turn exit was chosen as the

most navigation mark in use.

3 RESULTS

The entire collection of AIS data frames was stored

in an SQL database for easy management and ex-

traction. Data frames was selected according to area

and ordered by time and MMSI number. The image

registration routines and data processing was im-

plemented in MATLAB. The time-series was con-

verted to images with the hist3 function of Math-

works’ “statistics toolbox” for MATLAB, and

optimization of image similarity was handled by the

constrained optimization function fmincon from “op-

timization toolbox” with gradient descend search.

3.1 Separation of traffic

Traffic clustered in seven groups, in addition groups

consisting of one to five time-series was also pre-

sent, but these have lack the numbers required to

proclaim them as traffic-groups. Of the seven major

groups one group consisted of AIS position reports

of vessels at anchor in the harbor and was excluded

from further analysis. The number of time-series in

the other six groups, as well as the breakdown in di-

rectional groups, is seen in Table 1

Table 1. Distribution of time-series into groups of geometric

similarity

___________________________________________________

Group Total Direction 1 Direction 2

___________________________________________________

1 1017 443 573

2 809 436 373

3 76 16 60

4 552 240 311

5 17 4 13

6 11 2 9

___________________________________________________

The geometric group of time-series belonging to

the five first groups is seen in Figure 2-6. Time-

series group six is excluded since it is only a small

component intersecting with the areas northeast cor-

ner.

Figure 2: Position trace of time-series in group 1

Figure 3: Position trace of time-series in group 2

Figure 4: Position trace of time-series in group 3

307

Figure 5: Position trace of time-series in group 4

Figure 6: Position trace of time-series in group 5

For further analysis based on statistics, only the

three densest populated geometric groups should be

considered. This is due to the uncertainty associated

with statistical analysis of small populations and to

avoid drawing conclusions on a weak statistical

base.

3.2 Maneuver sequences and statistics

Maneuver sequences were identified and statistics

for traffic properties and maneuver parameters were

produced. The measured variables were fitted both

to the normal and skew-normal (Azzalini 1985)

probability distributions. The skew-normal distribu-

tion was introduced to compensate for expected

skewness in the data that could severely influenced

the accuracy of the normal fit. Statistic calculations

and fitting of distributions was performed with “R”

with the “MASS” and “SN” statistic libraries. The

median paths with turn section border points indicat-

ed are shown in Figure 7.

Due to space limitations a full treatment is only

possible for one direction in one of the sample

groups. The direction group 1 of sample group 4 is

analyzed further. The parameters of fitted skew

normal probability distributions of the traffic param-

eters are shown in Table 2.

Figure 7: Median traffic paths with turn sections indicated.

Table 2. Parameters for traffic statistics for group 4, direction 1

___________________________________________________

Section Type Variable Location Scale Shape

___________________________________________________

1* Turn k.ext [1/m] -5.20e-4 2.75e-3 -14.6

speed [kn] 5.5 3.42 3.05

2 Straight offset start [m] 19.7 37.3 -1.13

offset end [m] -37.5 64.6 0.85

course [rad] -0.90 0.08 1.89

speed [kn] 7.9 3.7 2.51

3 Turn k.ext [1/m]-4.21e-4 7.36e-4 -9.34

speed [kn] 8.3 4.15 3.12

4 Straight offset start [m] -100.4 155.0 1.74

offset end [m] -120.6 163.9 2.36

course [rad] 0.012 0.053 1.18

speed [kn] 8.5 4.30 3.48

___________________________________________________

* S

tart section inside harbor

From Table 2 it is possible to track the increase in

both the vessel speed and spread from median posi-

tion from the harbor area to the edge of the studied

area. The curvature of the turn sections shows that

the course-changing maneuver in section 3 can be

modeled as a turn-circle maneuver with a radius of

approximately 1.25 nautical miles.

The goodness of fit between the data from the

AIS time-series and the skew-normal probability

distribution function can be seen in Figures 8-10

where the empirical density function is plotted to-

gether with the fitted distribution functions.

Figure 8: Extreme value of curvature during turn.

308

Figure 9: Offset from median position at end of turn (start of

next section)

Figure 10: Median speed over turn section

3.3 Navigational aids

The identification of navigational aids made use of

the information of navigational markings in the area

as provided by the Norwegian hydrographic Service,

but excluded markings consisting of iron poles used

to mark shallows. This left only lighthouses and

light buoys. Identification showed good consistency

with only two to three objects contributing the ma-

jority of observed identifications. For the initiation

of the turn in section three seen in Table 2 the rela-

tive contributions are shown in Figure 11

Figure 11: Relative frequency of identified used navigational

aid

From Figure 11 it is apparent that the navigation-

al markings at the location “Nesjaflua” dominates as

the most probable navigational mark. The distribu-

tion of course angles between the two most used

markings is seen in Figure 12, the overlapping

notches in the plot indicates that there is no statisti-

cally significant difference between the median ap-

parent angle to the markings.

Figure 12: Distribution of apparent angle to landmark, box

width indicates sample size.

4 CONCLUSION

It has been demonstrated that image registration

techniques can provide an efficient and accurate so-

lution to the problem of shifting through large

amounts of position reports from AIS and prepare

them for analysis in groups. Image registration also

overcomes the problem of identification of turn ma-

neuvers in individual time-series. The group analysis

of the AIS position reports enables the identification

of statistical parameters for the traffic flow, as well

as of probable navigational marks for turn initiations

and turn radi.

REFERENCES

Aarsæther K. G. & Moan T. 2007. Combined Maneuvering

Simulations, AIS and full-mission simulations. In Adam

Weintrit (ed) Advances in marine navigation and safety of

sea transportation; proc 7. intern. symp. on navigation

Gdynia 20-22 June 2007

Azzalini A. 1985. A class of distributions which include the

normal ones. Scandinavian Journal of Statistics 12:171-178

Brown L. G. 1992. A survey of image registration techniques.

ACM computing Surveys 24:325-376

Gucma L. & Goryczko E. 2007. The implementation of oil

spill cost model in the southern Baltic Sea area to assess the

possible losses due to ship collisions. In Adam Weintrit

(ed) Advances in marine navigation and safety of sea trans-

portation; Proc 7. intern. symp. on navigation Gdynia 20-

22 June 2007

Hutchison, B.L.; Gray, D.L. & Mathai, T., 2003. Maneuvering

simulations - an application to waterway navigability.

Transactions of the Society of Naval Architects and Marine

Engineers Vol 111:485-516

Merrick J.R.W.; van Dorp J.R.; Blackford J.P.; Shaw G.L.;

Harrald J. & Mazzuchi T.A. 2003. A traffic density analysis

of proposed ferry service expansion in San Francisco Bay

using a maritime simulation model. Reliability Engineering

and System Safety. Vol 81(2):119-132

Vetterling W.T.; Press H. P.; Teukolsky S. A.; Flannery B. P.

2007. Numerical recipes – the art of scientific computing.

Cambridge university press

Zitová B. Flusser J. 2003. Image registration methods: a sur-

vey. Image and vision computing 21(11):977-1000