Вы находитесь на странице: 1из 6

Extended Grid Based Collision Avoidance

Considering COLREGs for Vessels


Michael Blaich Michael Rosenfelder Michael Schuster
Oliver Bittel Johannes Reuter

University of Applied Sciences Konstanz, Konstanz, 78462 Germany


(e-mail: mblaich@htwg-konstanz.de).

Abstract: In this work a grid based collision avoidance algorithm which considers the physical
constrains of a vessel is presented. For this purpose a new geometry neighbourhood is introduced
and explained in detail. The collision avoidance algorithm pays attention to the COLREGs and
provides a collision-free path. To find this path, Lees algorithm is used.
Keywords: Collision avoidance, Ship navigation, Path planning, Lees algorithm, COLREGs,
Optimal route, Raster grid
1. INTRODUCTION
In the last decades, vessel traffic and average cursing speed
have increased, thus the collision risk for vessels is rising.
Many Collision Avoidance (CA) algorithms are carried out
in maritime navigation research to decrease this collision
risk. Most research is carried out to develop supporting
systems for vessel navigators. However, most of these algorithms consider only other vessels but not the environment
e.g., the shore line or other static obstacles. Another way
to prevent collisions is to develop systems which can autonomously generate manoeuvres to avoid collisions. These
systems have to consider both, other vessels and static
obstacles. Nowadays, collision avoidance manoeuvres considering the local traffic and other obstacles are usually
subject to the navigators own reaction and judgement.
To adjust several avoidance manoeuvres, traffic rules are
required. Therefore, the only specific rules for maritime
navigation are the collision regulations (COLREGs 1 ). As
reported in Perrow (1984), 56% of the maritime collisions
are caused by the violation of the COLREGs. Thus, a CA
system which considers the COLREGs could be benefiting
for the vessel navigator.
There are several navigational advising systems to simplify
the decision finding process for the navigator, but they
do not autonomously avoid collisions. For example, the
Vessel Traffic Service (VTS), is a marine traffic monitoring
system established by harbour authorities to observe the
marine traffic. The idea of the system is a central unit,
receiving navigation information of all vessels and sending
the traffic situation to every connected vessel. Another
system is the Automatic Identification System (AIS). AIS
is an automated tracking system used on vessels and is
included into VTS for identifying and locating vessels by
electronically exchanging data with other nearby vessels
and VTS stations. The Automatic Radar Plotting Aid
(ARPA) is also used for tracking other objects. A radar
with ARPA capability can create tracks using radar contacts. The system calculates the tracked objects course,
1

The International Regulations for Preventing Collisions at Sea

speed and Closest Point of Approach (CPA). A possible collision with other vessels or shore lines can so be
recognised. One major drawback is, that these systems are
expensive and not suitable for small vessels like they are
used on the Lake Constance.
This work focuses on a CA for small vessels in inland
waters or ports where high traffic density and narrow environment leads to high collision risks. Thus, the close range
collision avoidance methods have become an important
subject. The developed algorithm can be used for a fully
autonomous collision avoidance, since it considers other
vessels, static obstacles like a jetty, a bridge or the shore
lines, and the COLREGs. It provides a collision free path
regarding the physical constrains of a vessel. This ensures
that the vessel can follow the provided path. The algorithm
calculates the CPA with other vessels or static obstacles
in a local area of up to 800 m. This range corresponds to
the radar sensor range used to detect other vessels.
In general, there are two possibilities to avoid collisions.
First, is to adjust the vessels velocity according to avoid
collisions with other vessels. The second is to dismiss
the current path and continue on an alternative course.
The presented Extended Grid Based Collision Avoidance
(EGBCA) approach focuses on planning an alternative
collision free path. But the speed altering for collision
avoidance could also be used by calculating different trajectories depending on different speed. The algorithm for
the alternative path calculation is based on the approach
of Szlapczynski (2005) and uses a raster grid. For the raster
grid, a variation of the maze-routing algorithm presented
by Lee (1961) is implemented to find a collision-free path
from the current position to a given goal. The mazerouting algorithm always finds the shortest path. The
drawback of this path is that it could include many turns
which lead to an unqualified vessel course. Szlapczynski
(2005) extended the maze-algorithm by introducing turn
penalties to achieve more suitable path with less turns
for vessel manoeuvres. Physical constrains like the turning
rate are not considered in Szlapczynski (2005), which limits the algorithm. In his approach, the vessel can reach all

research started with a fuzzy set approach published by


James (1986).

Fig. 1. The vessel Korona from the HTWG Konstanz used


for dimensioning and testing the CA algorithm on the
Lake Constance.
neighbour cells in the grid apart from the previously used
cell. To be able to turn a real vessel about 135 degree in
one cell, a large cell size is required. For our vessel Korona
shown in Figure 1, the cell size would be about 40 m. In
narrow environment like ports or channels, this cell size is
not suitable for CA. Therefore, the EGBCA extends the
grid by a cell size independent model of the vessel dynamic
which regards the physical constrains of the vessel.
To use the CA algorithm on real vessels, some further
extensions are required. First, the algorithm has to keep
the conditions of the COLREGs. This is realized by
adopting the ship domain of Goodwin et al. (1975) for the
other vessels. Further, the collision-free path should have
a tangential connection to the original path, in this work
called global path. This is necessary because the CA only
works in a local domain of 800 m but the global path can
be much longer. Therefore, the collision-free path requires
a tangential connection to the global path to get a smooth
changeover from the local collision-free path to the global
path.
This work is structerd in 5 sections. In section 2, similar approaches for maritime collision avoidance are introduced. Section 3 presents a new grid based approach to
overcome the limitations of the previous ones. The experimental results are shown in section 4. Finally, section 5
concludes this work with some future prospects.
2. RELATED WORK
Collision avoidance of vessels has been first discussed by
Sharpey-Schafer (1955). The discussion focused on the behaviour of marine traffic in general and optimal strategies
for evasive manoeuvres in close range encounters. This
work concentrates on general issues and provides some
suggestions how to act in a possible collision situation.
Collision avoidance research for vessels, nearly stopped
for the next decades after this discussion. Because of,
increasing traffic and collisions on sea in the last two
decades, research focused on theoretical and practical developments of collision avoidance systems. The intensive

Nowadays, many different approaches are carried out for


maritime collision avoidance. An overview of the algorithms developed in the last decades is given in Statheros
et al. (2007) and Tam et al. (2009). Referring to these
publications, the CA algorithms can by divided into two
groups, the deterministic and the heuristic approaches.
Deterministic approaches use mathematical models and
algorithms to find an optimal solution. Depending on
the problem complexity and the large search space, the
computing time of these algorithms is high. Thus, most of
these approaches are not suitable for a real CA system. To
reduce this computing time, heuristic approaches search
for an acceptable solution in a subspace. Therefore, often
soft-computing techniques based on Artificial Intelligence
(AI) are adopted. Some typical soft-computing methods
used for CA algorithms are evolutionary algorithms, fuzzy
logic, neural networks, expert systems and combinations of
these as hybrid systems. Referring to Tam et al. (2009), the
most promising algorithm is the evolutionary algorithm
published by Smierzchalski (1999). But for this work the
focus is on a deterministic approach. Tam et al. (2009)
claim that the approach of Szlapczynski (2005) is the most
practical and the most efficient one. The approach is based
on Chang et al. (2003) which uses a raster grid and the
maze-routing algorithm.
Approaches with raster grid realize the search by iterating
over predefined steps and guarantee a solution if there is
one. A raster grid is an extension of an occupancy grid
but stores more then one value per cell. The grids are
used to discretise the search space for a faster processing.
These approaches have become more popular in the last
decade. The reason is the increasing storage space and
computer performance. Search algorithms like the mazerouting algorithm used in Chang et al. (2003) find an
optimal path in these grids. The grid raster of Chang
et al. (2003) stores in each cell if it is sea or land, if the
cell is occupied by an other vessel and the arrival time of
the own vessel. The maze-routing algorithm always finds
the shortest path. This does not have to be the optimal
path for vessel navigation if it consist of many turns.
For vessel navigation, a straight path with a minimum
of turns is mostly preferred by navigators. To solve this
problem, Szlapczynski (2005) presented a modified version
of this approach and added turn penalties to each cell.
This results in a path with a minimum of turns which is
more suitable for vessel navigation. Additionally, he added
the possibility to reduce the own vessels speed to avoid a
collision. This speed altering is only used if no collision-free
path exists for the current speed.
3. EXTENDED GRID BASED COLLISION
AVOIDANCE
The EGBCA is based on the approach of Szlapczynski
(2005) who extended the approach of Chang et al. (2003)
by turn-penalties. Both approaches will not be further explained, it is assumed that the reader is familiar with these.
The main improvements of the EGBCA regarding to the
previous approaches are a new neighbourhood geometry
called T-neighbourhood and the tangential connection to

the global path. The T-neighbourhood is used to consider


the vessels physical constraints. For this purpose, a new
storing method for the orientation of the vessel in the
grid is required which leads to a new calculation of the
turn penalties. This extension provides the possibility to
use a smaller cell size which improves the accuracy and
the consideration of the COLREGs because of a better
approximation of the ship domain.
Normally the vessel follows a predefined global path. This
path is given be an autopilot system or by the vessels
navigator and is not checked for collisions. The EGBCA
checks if collisions occur in a subsection of the path. This
subsection is also called local area. The size of the local
area is adopted to the sensors range for detecting other
vessels. For this work a grid size of 400 cells with a cell
size of 2 m is used. Only considering a subsection of the
global path leads to the problem that the vessel has to
return to the global path after a collision is avoided.
To get a drivable path, the reconnection to the global
path should be smooth. Therefore, the EGBCA uses a
tangential connection between the local collision-free and
the global path.
A further extension for the collision avoidance process is
the providing of supporting points of a trajectory instead
of list of visited cells as done in the previous approaches. To
find the supporting points of the visited cells, the Douglas
and Peucker (1973) algorithm is used. As presented in
Schuster et al. (2012), supporting points can be directly
used for navigation systems to control the vessel.
3.1 Considering Vessels physical constraints
Each vessel has its own dynamic characteristics. For CA,
the most important characteristic is the maximal turning
rate which defines the turning circle. This turning rate is
given in degrees per travelled meter. If the CA does not
consider this turning rate, the vessel may not be able to
navigate on the planed path. For this work the physical
constrains of the Korona are used, with a turning rate
of 0.66667 /m. This means for a track length of 10 m ,the
Korona can change the orientation of about 6.6667 . To regard this performance, the T-neighbourhood is introduced.
An overview of the traditional geometry neighbourhood is
given in Chang et al. (2003). Contrary to the traditional
geometry neighbourhoods the T-neighbourhood only considers the cells reachable by the vessel even for a small
cell size. This leads to a neighbourhoods geometry which
looks like the character T. For this work, the vertical line
is called body and the horizontal line is called head. In
this work, the T-neighbourhood is normalized to a head
of three cells. The physical constrains can be applied by
modify the length of the body. Using the Korona and a cell
size of 2 m e.g., leads to a T-neighbourhood with a head
of three cells and a body of four cells which is shown in
Figure 2.
Using the T-neighbourhood requires a precise orientation
which is not reachable with the discretised orientations
used by traditional geometry neighbourhoods. They reach
a maximum step size for the orientation of 11.25 for a
16-geometry neighbourhood. This is not suitable for the
T-neighbourhood. To solve this problem, the orientation
of the vessel is stored in an undiscretised form in each

Fig. 2. Example of the T-neighbourhood for the boat


Korona and a cell size of 2 m.

Fig. 3. The domain presented by Goodwin et al. (1975)


and its basic representation (if a ship is considered as
one cell) in the grid raster
cell. Therefore, the EGBCA only discretise the position of
the vessel including the latitude and longitude coordinates.
But the incoming orientation for each cell is stored as an
undiscretised value. Using these undiscretised orientation
values the turn penalties have to be calculated between
the incoming and outgoing orientation. This leads to new
turn penalties. To adopt this turn penalties to the cell size
of the grid, a scaling factor is added.
3.2 Considering the COLREGs
As a benefit of the T-neighbourhood a raster grid with
a small cell size can be used. This provides a precise
realisation of the ship domain presented by Goodwin et al.
(1975). This approach obtaines a non-symmetric ship domain which is divided into three sectors as shown in Figure
3. For each sector a different dimension is used to model
the collision risk for the direction of this sector. Using this
ship domain for other vessels and Lees algorithm to find
the collision-free path enabled the consideration of the
COLREGs since Lees algorithm finds the shortest path
to pass this domain. Therefore, the path on the stern side
is shorter than on the bow side and on the port side it is
shorter then on the starboard side. In a heading situation
e.g. , this leads to avoidance manoeuvre for the own vessel
to the starboard side to pass the other vessel on the port
side. This is conform to the COLREGs. Detailed tests with
considering the COLREGs by the EGBCA are presented
in section 4.
3.3 Tangential Connection to the Global Path
In vessel navigation, typically a global path is planned and
a CA algorithm checks in a local area if the subsection of
the path is collision-free. A scenario with a global path

5
0
5

20

10

10

20

30

40

Fig. 6. Overtaking scenario.


4.3 Crossing Port Side
The crossing port side scenario shows the crossing by
another vessel from port side. The EGBCA plans a path
to pass the other vessel on stern side as shown in Figure 7.
Fig. 4. Consideration of tangential connection to the global
path.
45

and a local area for the CA is shown in Figure 4. The


blue line is the global path and the vessel is shown in the
middle of the red marked window which defines the local
area. The shore line and the pier are marked orange and
the vessel is highlighted in red. If the CA generates a new
collision-free path, it has to be connected tangential to
the global path. To consider this tangential connection,
two new virtual circular obstacles are added to the grid.
This can be seen at the right part of Figure 4. These
obstacles are called funnel, based on their functionality.
The diameter of these obstacles corresponds to the turn
circle of the vessel to consider this physical constraint.
These obstacles are created by a wave-front propagation
algorithm.

40
35
30
25
20
15
10

4. EXPERIMENTAL RESULTS

To analyse the performance of the EGBCA and the


consideration of the COLREGs, several test scenarios are
carried out. These scenarios have been verified by Perera
(2010) with his fuzzy logic approach. For all tests, two
vessels with different course conditions are used. The own
vessel has a speed of 1m/s and the other vessel 2m/s. In
all the scenarios, collisions have been avoided confirm to
guidelines of the COLREGs.

0
5
10

10

20

30

40

Fig. 7. Crossing port side scenario.


4.4 Crossing Port Bow Side

4.1 Heading
In the heading scenario the other vessel drives straight
towards the own vessel. The result of the EGBCA is a path
passing the other vessel parallel on port side as shown in
Figure 5.

In this scenario, the other vessel drives towards the own


vessel from port bow side. The EGBCA plans a path
to pass the other vessel on stern side. This is shown in
Figure 8.
35

5
30

0
5

25
0

10

20

30

40

50

60

Fig. 5. Heading scenario.

20

15

4.2 Overtaking
If the own vessel is overtaken by another vessel, normally
the own vessel stays on the course. But if the other vessel
has a higher priority as the own vessel, the own vessel
has to dismiss his path. The overtaking scenario simulates
an overtaking by a vessel with higher priority. For this
scenario the EGBCA dismisses its path and successfully
avoids a collision as shown in Figure 6.

10

10

20

30

Fig. 8. Crossing port bow side scenario.

40

4.5 Crossing Port Stern Side

In this scenario, the other vessel drives towards the own


vessel from port stern side. The EGBCA plans a path
which first goes parallel to the other vessel and then passes
on stern side. This is shown in Figure 9.

10

35

15

30

20

25

25

20

30

15

35

10

10

20

30

40

50

Fig. 11. Crossing starboard bow side scenario.

4.8 Crossing Starboard Stern Side

10

10

20

30

40

Fig. 9. Crossing port stern side scenario.

In this scenario, the other vessel drives towards the own


vessel from starboard stern side. The EGBCA plans a path
which first goes parallel to the other vessel and then passes
on stern side. This is shown in Figure 12.
10

4.6 Crossing Starboard Side


5

In the crossing starboard side scenario, the other vessel


drives towards the own vessel from starboard side. The
EGBCA plans a path to pass the other vessel on stern
side as shown in Figure 10.

10

5
15

0
20

5
25

10

30

15

10

20

30

40

Fig. 12. Crossing starboard stern side scenario.

20

25

5. CONCLUSION
30

35

40

45

10

20

30

40

Fig. 10. Crossing starboard side scenario.


4.7 Crossing Starboard Bow Side
In this scenario, the other vessel drives towards the own
vessel from starboard bow side. The EGBCA plans a path
to pass the other vessel on stern side. This is shown in
Figure 11.

This work presented a grid based collision avoidance algorithm. The algorithm uses a special T-neighbourhood
for the grid search algorithm to consider the physical
constraints of the vessel. This new T-neighbourhood offers
the possibility to use a smaller cell size. Some of the
deficiencies reported by Tam et al. (2009) for the approach
of Szlapczynski (2005) are eliminated by this work. The
extendability to other objectives is guaranteed by the
variability of the cell size. Smaller cell sizes also lead to an
improved consideration of the COLREGs. The confirmation of the produced collision avoidance manoeuvres with
the COLREGs is proved in several tests.
For future work the T-neighbourhood could be modeled
dynamically and asymmetric to regard external influence
e.g., wind and river current.

REFERENCES
Chang, K.Y., Jan, G.E., and Parberry, I. (2003). A Method
for Searching Optimal Routes with Collision Avoidance
on Raster Charts.
Douglas, D. and Peucker, T. (1973). Algorithms for the
reduction of the number of points required to represent
a digitized line or its caricature. Cartographica: The
International Journal for Geographic Information and
Geovisualization, 10(2), 112122.
Goodwin, E., Polytechnic, L., and for National Academic Awards, C. (1975). A statistical study of ship
domains. Cambridge Univ Press.
James, M. (1986). Modelling the decision process in
computer simulation of ship navigation. Journal of
Navigation, 39(01), 3248.
Lee, C. (1961). An algorithm for path connections and its
applications. Electronic Computers, IRE Transactions
on, 30, 13891401.
Perera, L.P. (2010). Autonomous guidance and navigation
based on the COLREGs rules and regulations of collision
avoidance . Design, (1999), 205216.
Perrow, C. (1984). Normal Accidents: Living With HighRisk Technologies. Basic Books, Inc., New York, USA.
Schuster, M., Reuter, J., and Blaich, M. (2012). Integration von Navigationsalgorithmen in Regelkreise am
Beispiel autonomer Systeme auf dem Wasser. Karlsruher Geowissenschaftliche Schriften, Rehie B - Vermessung und Photogrammetrie.
Sharpey-Schafer, J.M. (1955). Collision at Sea. Journal of
Navigation, 8(03), 261280.
Smierzchalski, R. (1999). Evolutionary trajectory planning
of ships in navigation traffic areas. Journal of marine
science and technology, 4(1), 16.
Statheros, T., Howells, G., and Maier, K.M. (2007). Autonomous Ship Collision Avoidance Navigation Concepts, Technologies and Techniques.
Szlapczynski, R. (2005). A New Method of Ship Routing
on Raster Grids, with Turn Penalties and Collision
Avoidance.
Tam, C., Bucknall, R., and Greig, A. (2009). Review
of Collision Avoidance and Path Planning Methods for
Ships in Close Range Encounters.

Вам также может понравиться