ControlEng. Practice,Vol. 4, No. 3, pp.

393-400, 1996
Copyright 1996 ElsevierScienceLtd
Printedin Great Britain. All rightsreserved
0967-0661196 $15.00 + 0.00

Pergamon
PII:S0967-0661(96)00017-2

UNDERWATER SONAR RANGE SENSING AND 3D IMAGE

FORMATION
P.G. Auran* and O. Silven**
*Department of Engineering Cybernetics, Universityof Trondheim, The Norwegian Institute of Technology,
N-7034 Trondheim, Norway (Per.Gunnar.Auran@itk.unit.no)
**Department of Electrical Engineering, Universityof Oulu, SF-90570 Oulu, Finland

(Received August 1995; in final form November 1995)

A b s t r a c t . A 3D spatial grid for exploiting the range and direction information inherent in sonar
range data is presented. Special attention is given to the real time performance of the representation,
and encouraging results have been demonstrated by computer simulations and sonar experiments.

K e y W o r d s . Underwater computer vision, 3D sonar data representation.

1. INTRODUCTION

grid framework could be used for integration of

multisensor data (for a reference to this principle, see (Durrant-Whyte, 1988)). This model is
suitable for real-time obstacle avoidance, mobile
robot mapping and navigation, and it supports
robust sensor interpretation methods, incremental discovery procedures and the explicit handling
of uncertainty (Elfes, 1990).

Intelligent vehicles will play a major role in future underwater missions, and the independent
and meaningful interpretation of sensor data is a
principal prerequisite for accomplishing high-level
autonomous underwater vehicle (AUV) missions
(Brutzmann et al., 1992). The interpretation of
sonar range data is specially important, as the
sonar is one of the most useful sensors for underwater geometric exploration. Much work has already been reported on sonar vision processing;
some examples could be found in (Brutzmann,
1994; Elfes, 1986; Leonard et al., 1992; Mandelbaum and Mintz, 1994; Mathisen, 1993; Matthies
and Elfes, 1988; Moran et al., 1993; Smith et al.,
1987). Most of the work is, however, done in
two dimensions, i.e. it addresses the construction
of 2D cross-sectional geometrical models or environment maps. For the purpose of environmental modelling by using line and arch primitives
(Mathisen, 1993), it is assumed that the objects
of interest have a constant cross-sectional geometry. This assumption is perhaps more suitable for
controllable scenes of man-made objects than for
a "real-world" operating AUV. The AUV's perception process must be able to cope with sonar
input describing arbitrary shapes.

A major part of the research in sonar-based modelling is concerned with building 2D-maps using
ultrasonic (and stereo vision) sensors operating in
air, i.e. the issue is obstacle avoidance and path
planning of mobile robots operating in "officelike" environments. With the operation of a six
degrees-of-freedom AUV in mind, some remarks
ought to be made:
1. The ultrasonic range transducers have different properties compared to marine sonars,
i.e. broader beam and shorter range. Sensor range with respect to vehicle size is not
comparable with the AUV case, which will
affect the objectives of sonar perception and
navigation.
2. Although well-known, the significantly different characteristics of sound propagation in
air compared to water should be pin-pointed.
Effects like signal damping, background noise
and temperature gradients (Ol'shevskii, 1978;
Urick, 1983) are not recognised by experiments in air. The conclusions drawn from
such experiments will be limited for the AUV-

Of special interest is the research presenting the

occupancy grid framework (Elfes, 1986, 1990,
1991). The occupancy grid is a spatial map of
the environment, consisting of small cells which
are either "empty" or "occupied". Each cell's
state is estimated in a Bayesian stochastic approach. Elfes (1990) concludes that the occupancy

case.

3. The proposed Bayesian cell-updating approach mentioned in (Elfes, 1990) demands

an assumption of independent cell states to
393

394

P.G. Auran and O. Silven

be feasible. It also constitutes an extensive cell-updating scheme, as large amounts

of cells must be updated when new information enters the occupancy grid. In a real-time
perspective this is only feasible with a limited
number of cells, and in 3D these remarks are
of vital importance for the
of the
occupancy grid approach.
4. In the proposed 2D occupancy grids, the cell
size is quite coarse (comparable to sensor resolution at maximum range), and also fixed,
e.g. not exploiting the resolution properties
of the sensor (range accuracy varies with distance).
If the map is not intended for
extraoting geometric properties, this may not
be of significant importance.

feasibility

Traditionally, sonar data have been presented as

2D radar-like plots (range and bearing) meant for
human interpretation. But as sonar range data
are by their nature three-dimensional, they should
be interpreted as such by the AUV. Taking into
account the plane of sweep (for a narrow-band
sweep sonar) or the "beam-plane" (for a multibeam sonar), the bearing gives information in two
dimensions, i.e. the two angles in a spherical coordinate system (see Fig.l).

tive vision (Bajcsy, 1988), and the ideas presented

here are highly motivated by the occupancy grid
approach. A 3D occupancy grid is used to
actively sensed sonar range data.
This grid could be interpreted as a
of sensor information, useful for accomplishing the
AUV's current task.

effeclocal 319 map

tively organise

2. MOTIVATION
The main objective of the grid will be as a firstorder modelling process for AUV-relative navigation and world sensing. By this is meant how to
perform path planning and inspection, using sonar
ranging from the AUV's point of view. One actual
problem could be to plan navigation for getting a
better view of the object, possibly utilising other
sensors as well (for example using a video camera
at short ranges). Assuming the AUV to operate
in "exploration mode", the grid is a first-order
model of what the AUV senses from its current
position, integrated with what it has sensed
to this position. The 3D grid is therefore
AUV-centered.

its way

on

Sealevel

SOb

phi

Proje~tmnin

:~'::~:.:!!::"..:~:~!::::::i:~:'........
:":".::::::::_:~::.!!!!:~i;:'~!!!!::?"
Fig. 2. The 3D grid at the AUV's core for world perception.
Fig. 1. The AUV coordinate frame and the spherical
angles.

In this work the sonar is regarded as a range sensor with certain geometric and uncertainty features. More detailed modelling of underwater
echo-ranging requires complicated solutions of the
wave equation in a non-homogeneous medium,
which is outside the scope of this paper. The main
objectives here are range data organisation, uncertainty modelling and interpretation of stationary
objects for purposes like environmental mapping,
obstacle avoidance, and path planning. A spatial
framework for the exploitation and active use of
the 3D sonar range data is presented. An important aspect of this approach is the concept of ac-

At this point it should be noted that the grid is

not meant to be the only world model, but a fast
organisation of raw input data of current interest.
The AUV will depend on multisensor input to be
able to execute its task satisfactorily. The 3D grid
should therefore be seen in this perspective and regarded as a core of organised sonar data available
for several processes in the AUV's expert system,
as outlined in Fig.2. It could be interfaced with
a "navigator" process including a vessel dynamics model, as found in (Fossen, 1991), or with a
bathymetric map-positioning system with an inertial navigation system as presented in (Bergem,
1993). A control process for sonar scan directions (sweep planes and beam configuration) could
also be included in the expert system. Further-

Underwater Sonar Range Sensing and 3D Image Formation

more, a "modelling" process is suggested for extracting high-level geometric information, possibly also utilising object recognition by using an
object database. All these processes would run
in parallel and interact in a complex AUV expert
system, similar to the ideas mentioned in (Brutzmann et al., 1992).

y4-

395

. .~.

.~ 0

,+
" -t-

'-t%

2.1. Some Features of 2D Echo Ranging

A great deal of research has been done on 2D geometric modelling, and it is therefore natural to
summarise some of the features of 2D sonar ranging:
* Regions of constant depth (RCDs) are an important feature, useful for separating echoes
from different objects, and also for filtering
multipath echoes (Leonard et al., 1992).
* Tangential clusters (Mandelbaum and Mintz,
1994), or echoes with similar tangent direction, could be used to detect planar surfaces
and corners. This principle is suggested for
3D, but experiments are in 2D.
, In (Mathisen, 1993) a methodology is based
on least-squares fitting of line and arch primitives, and then producing hypotheses of objects or basic elements. It includes a parameter called "accounted for", a factor equal to
the number of echoes each object has produced. The amount of echoes, and also their
repetitions, indicate strong target features.
For fixed objects, echo stationarity is another
important feature.
The fact that sonar echoes near corners
poorly describe their location should be
acknowledged (Brutzmann et al., 1992;
Mathisen, 1993).
When the environment includes machinery,
for example in an underwater oil-production
scenario, passive sensing could be used to detect these objects. Rotating machinery could,
for example, be identified using its resonant
frequencies (Maksym et al., 1983). However,
this paper addresses active and controlled
sensing, and passive sonar sensing will not
be further considered here.

.4.._~2
~ ............

..-b

--I-

Fig. 3. 2D geometric model and vehicle path.

metric environment model consisting of lines and
circles. Inside this model a path is stipulated,
with '+' signs indicating the positions where sonar
range data are obtained (360 scan, limited sonar
range). Noisy echoes from two of the edges, obtained at the first position in the vehicle path,
are given in the figure. Assuming that the vehicle advances to the next position in the path,
it will also receive echoes from a third edge (second lower edge on the parallelogram). The vehicle
may now form a corner hypothesis, for example
by a parametric regression scheme (Brutzmann et
al., 1992). At the next two positions it is possible to verify the corner hypothesis by observing
that the echoes of the first edge disappear as one
moves round the object. In this way the dynamic
behaviour of echoes could be used as "items of evidence". If the vehicle movement and sonar scan
directions are controlled to search for and exploit
such events, the sensing will be active. The underlying echo representation could support this idea
and make it an integrated feature of the framework. Note also that events of this type could
provide useful position information for increasing
the navigation accuracy (assuming a priori knowledge of the scene). This is specially interesting if
the AUV is equipped with an inertial navigation
system where position errors increase with time
(gyro drift problems).

3.2. Grid Representation

3. THE 3D OCCUPANCY GRID
3.1. Active Sonar Sensing
By active sensing it is understood that the sonar
range data should be acquired in a controlled manner. Less sensing should be spent in areas of little interest, whereas sensing of objects and their
characteristic features should be more thorough
(Bajcsy, 1988; Blidberg, 1987; Elfes, 1990). The
active sensing approach is illustrated by the 2D
sketch in Fig.3. This figure illustrates a 2D geo-

The grid is a 3D sonar range image of the AUV's

close environment, i.e. the environment within the
sonar range represented by a dynamical window of
the world. Figure 4 illustrates a spherical versus
a rectangular slice of such a window. Considering
the characteristics of sonar data, it is clear that
this image should have certain properties:
1. Sonar range echoes must be organised in a
way that preserves the range and direction
information, i.e. it should be easy and fast to

396

P.G. Auran and O. Silven

*

(a) Spherical

(b) Rectangular

Fig. 4. Spherical versus rectangular grid organisation.

check whether movement in a given direction

is feasible or blocked.
2. The grid should reflect the accuracy of the
sonar range readings, i.e. the occupancy cells
should not be too large. This is specially
important for a modelling process using the
grid as input data for extracting geometric information. Important sensor characteristics
should be reflected in the data representation.
3. "Curse of dimensionality": In a fixed-size 3D
rectangular array, the preservation of sonar
accuracy would make heavy demands on computer memory, and grid updating would be
time-consuming. In contrast to the approach
in (Elfes, 1986, 1990), one must focus on
those cells that would be affected due to
nearby echoes, and concentrate the processing on those (Blidberg, 1987). As sonar
echoes in nature are sparse, this is a feasible
way of realizing a 3D occupancy grid, allowing reasonable updating rates and computer
memory demands.
4. Level of detail versus range: Moravec (1986)
suggests that objects at increasing range
should be represented with less detail to get
better real-time responses from algorithms
using the 2D occupancy grid. Distant cells of
the square grid were blurred to achieve this
effect.

60 beams
150 degrees sector width
2.5 degrees beam spacing
Accuracy 0.3% of water depth (surface vessel
mounted sonar) or maximum 15cm. resolution.

In a spherical grid it is quite feasible to let cell geometry reflect these properties. For short ranges,
the radial cell size could be equal to the maximum
resolution, whereas further away from the AUV,
the radial cell size increases as a geometric series.
As such, the grid could be viewed as spherical
layers of increasing thickness. For a fixed maximal range, this approach will still give a fixed
number of cells in the radial direction. Simulations have indicated that such a grid is feasible,
even in a real-time perspective (grid update rate
keeps up with sensor data rate). When implementing this grid, only the theta and phi angles
(standard spherical coordinate definitions, refer to
Fig.l) need to be quantified into "buckets", allowing flexibility in the range parameter. It would
also be useful to be able to store grids with different resolutions, specially when searching for echo
clusters. This idea is similar to the scale-space
principle often used in image processing.

THE~TA ANGLE CELL QUANTIFICATION

.>

I I '.'

I
RADIAl. CELL QUANTI HCAT1ON ~>

Fig. 5. 3D spherical grid data structure.

3:3. Echo Organisation
On the basis of the previous discussion, it is natural to choose a spherical coordinate frame for
the grid. This reflects the properties of the sonar
sensor in a very direct way, i.e. beam orientation and range readings. In this framework it is
straightforward to allocate a sonar echo directly
to the corresponding place in 3D space. Furthermore, it allows the cell geometry to reflect the
sonar geometric properties. As an example, some
of the properties of the Simrad EM 1000 multibeam sonar working in shallow mode are quoted
below (Simrad, 1995):

Figure 5 shows a feasible organisation of the grid

data structure. Grid cells are organised by two
doubly linked lists, one ordered by beam indexes
(the two spherical angles) and the other by range.
For each beam there is a list of cells, sorted by
distance. Echoes of short range will arrive at
the AUV before the more distant ones, and there
should be little overhead in maintaining this order in the representation. The different image
processing algorithms, for example 3D connected
components analysis, can be implemented more
efficiently in a well-organised grid. To allow fast
"look-up" of any beam direction, two hash-tables

397

Underwater Sonar Range Sensing and 3D Image Formation

are included in the structure. Note that the

phi table only exists for those theta quantifications where echoes were found, minimising memory storage and allowing fast checking of "empty"
theta and phi directions. Additionally, a binary
table is needed to separate empty beams from
space directions that have not yet been investigated (or "pinged"). The links between beam cells
is ordered as the hash-tables indicate: increasing
phi for each theta angle quantification. This doubly linked list allows fast data processing of only
the cells which have actually been stored, without traversing all the elements of the hash-tables.
For each cell, several features are stored: radial
limits, probability state, total number of echoes
in this cell, and time elapsed since the first and
last echo. A parameter called "flag" in the figure
is used to reflect the dynamic properties of a cell.
This parameter is set to different values according
to whether the cell appeared or disappeared in the
last scan, or if it was recently updated. It also includes a "lock" bit which can be used to freeze the
cell state. Furthermore, another list between cells
at similar radial distance could be implemented
(utilising the ThisRad pointers), a list specially
useful for RCD (Leonard et al., 1992) or tangential cluster (Mandelbaum and Mintz, 1994) based
algorithms.
When objects are investigated from different positions and viewing angles, the spatial echo uncertainty can be reduced (Elfes, 1986). A drawback
of the spherical representation is that the cell geometry also changes with the phi angle, and recently a sub-cell representation has been included
to implement a more general measure of uncertainty. Each cell stores an echo interpolation point
contained in an uncertainty ball. Vessel motion errors, and reduced spatial uncertainty due to multiple echoes within the cell volume, can then be
represented with better accuracy.
The orientation of the grid is another interesting question. It would be advantageous to keep
the orientation fixed when referring to the location of a priori known objects using some sort
of area maps. By using information from other
parts of the on-board AUV expert system, such
as the inertial navigation and dynamics control
systems (including current and disturbance measurements), the grid could (as an example) be
fixed with the x-axis in the East direction and z
positive upwards. Another possibility is to define
a "course-sensitive grid" using the current course
as the y-direction (refer to Fig. 1). The symmetric properties of the spherical grid allow flexibility in orientation. Notice how straightforward
cells could be rotated by only updating the theta
and phi pointers. This grid is implemented as a

dynamic pointer structure in the C programming

language, allowing fast creation and deletion of
multiple grids. The presented structure is, however, not intended to be final or complete. The
important issue here is the coupling of echoes and
echo features with their respective 3D spatial locations.

4. SIMULATIONS AND EXPERIMENTS

4.1. The Sonar Sensor Model
Initially, some simulations were performed to illustrate the methodology. A 2D sonar simulation
model, based on experiments from a Mesotech
971 sonar (Mathisen, 1993), was used to generate sonar-like data. This model is basically a
2D ray-tracing program that accepts an arbitrary
input geometry consisting of arc and line segments. The rays are under-sampled to a corresponding sonar beamwidth (2.5 degrees), and
Gaussian noise is added according to a model similar to the one used in (Elfes, 1990):

p((r,O)l(R,O)) =

2~rdRdO X

(1)

exp
This 2D polar Gaussian probability model expresses the probability of an echo at one point
(r, 0), given a sonar range reading at another one
(R,O), and assuming independent errors in radial and angular coordinates. The standard deviations (dR and dO) are given by the characteristics of the sonar sensor. Note that dR is a function of range, typically given as a percentage of
distance. For 3D simulation the model was used
to produce sonar-like echoes from cross-sections
at different z-coordinates. When mapping sonar
echoes to grid cells, the probability model in equation (1) is extended with another angular parameter of equal size (assuming conical beams). This
model is used to map sonar range readings into a
grid cell and update the state parameter. Initially,
a simple scheme using cell midpoints and maximal responses was implemented. More sophisticated algorithms have recently been investigated
using real sonar data as input, and specially the
inclusion of relative time between echoes allows a
more realistic modelling of cell states (Auran and
Malvig, 1996a).

Not only cells with actual sonar responses in them

should be stored in the grid. A 3D "uncertainty
neighbourhood", centered at the physical echo
would be affected, i.e. some sort of relaxation
scheme will be useful. Some experiments with 3D

398

P.G. Auran and O. Silven

masks related to the Gaussian noise model have

been done. The size of these masks is set proportional to the standard deviations in the probability model using varying scale factors. Because of
the sparse nature and relatively low data-rate of
the sonar, this idea could be used by the interpretation process to form connected clusters of cells,
i.e. to group echoes into clusters corresponding
to objects. This approach has recently been investigated for 3D connected components analysis
(Auran and Malvig, 1996b). Note, however, that
the 3D mask sizes must be chosen carefully in order not to be too resource-demanding.

-200

700
600
50O

+" 2

4.2. Simulation and Visualisation

+'3

20O
x

Simulation and visualisation are useful and necessary tools for AUV design (Brutzmann, 1994).
At the Department of Engineering Cybernetics,
there has been a similar research project called
M O B A T E L (Balchen and Yin, 1994), which also
emphasises the use of powerful graphical workstations. Graphical visualisation is specially important when working in 3D. A visualisation tool has
been implemented, using Matlab to inspect the
output of different algorithms implemented in C.
This tool allows inspection from arbitrary viewing
directions, including zooming. The grid is presented as a 3D grey-level image, with the grey
level reflecting the cell state (darker cells represents more reliable echoes). An example of such
a 3D sonar image is given in Fig.6. Figure 6(a)
illustrates the 2D cross-section and the 3D path
within this virtual geometry. The indicated path
is constructed by adding a noisy sinusoidal movement in the z direction to a 2D path in the xyplane. The first ten "sonar scan locations" along
the path are enumerated. The second part of the
figure, Fig.6(b), is the resulting grid in position 10
as seen from above, keeping track of sonar scans
taken along the path. The four outer corner points
of each cell are also plotted, to make small cells
visible. The third picture, Fig.6(c), is a 3D view of
the same grid, illustrating its real structure. The
dotted lines are spaced 30 degrees apart in the xyplane centered at the current position. Note that
cell sizes vary with distance, in both radial and
angular directions.

(a) Geometric model

30~

20~

100

-100

eg

-200

.300
-200

-100

100

200

300

W,~"
400

/ '
500

(b) 2D crossection

200,
150,
100,

50,
0,

East

-5D,
-100.
-150,
-200

200

4.3. Experiments
A specialised hardware circuit has recently been
designed to sample 3D echoes from a WesMar
SS150 sweep sonar. This is a mechanical sweep
sonar with a sound-cone of 6.5 6.5 and a frequency of 160kHz (WesMar, 1970). The sonar
is mounted on the research vessel "Balchazar",
a surface vessel belonging to the Department of
Engineering Cybernetics. The sonar b e a m d i r e c -

100

Y
x

(c) 3D grid
Fig. 6. Path in geometric model and resulting 3D grid
(in model coordinates).

399

Underwater Sonar Range Sensing and 3D Image Formation

31HI

270

2411'

ii
2

'30

120
~

Is0
90

I$0

(a) Scan angle

(b) Tilt angle

8akhazar

Fig. 7. Definition of sonar scan and tilt angles with

one sonar beam indicated (angles in degrees).

&

(a) Environment crossection

tion is servo controllable in the tilt and scan angles, and DC signals are available for measurement of the beam direction (voltages are linear
in angle). As can be seen from Fig.7, these angles are directly related to the spherical angles
theta and phi. The hardware circuit samples
the scan and tilt angle DC signals of each beam
and the backscattered echoes along the soundcone centered at this direction. Only the echoes
above a certain amplitude level are sampled (analog thresholding)~ and thercfore each echo sample
consists of a time/amplitude pair. The threshold
level is set by a D / A converter (software controllable). An illustration of one of the experiments
performed in the Trondheim harbour is given in
Fig.8. The vessel motion is minimised during the
experiment. A sketch of the environment and vessel position is given in Fig.8(a) and the resulting
prefiltered echoes are shown in Fig.8(b). A noticeable echo gap for zero scan angle is visible in
the second figure, which most likely is due to a
small change in the orientation of the vessel. Figure 8(c) shows a view of some of the resulting 3D
echo clusters. This figure also illustrates the use
of echo relaxation; notice that the two major echo
clusters in front of the vessel are connected to a
single one. Further experiments and a description
of the sonar prefiltering and grid-updating processes are found in (Auran and Malvig, 1996a).

S o n a r Scan

02O
330 . . . . . . . . . . . .
..... '
..-'"
16
..
.
. ....... ! .....

300 / "

"

,/'.

..

30
.,:.

12

. ""

"'".." . . . . . . .

"+

".

". 60

2 7 0 i ...... i . . . . i ...... " ...... : . ~ ; o . . . . . . . . . . . . . .

=.
-.

.'':

..,

"4
~
. ....;..

ve

24o .

..

:.

" .....

~.. ~.

-.

.,
210

: ...........

::

90

12o

...
.....

......

150

180

(b) Prefiltered echo crossection

3,
2,
1,
~'o,

N-l,
-2,
-3,

5. CONCLUSION
A data structure for real time 3D sonar management has been presented. Experiments indicate
that the algorithms will keep up with a data rate
more than 60 times that of the WesMar SS150
sonar when implemented on a ordinary Sun Sparc
2 workstation (Auran and Malvig, 1996b). This
performance includes grid maintenance (registration, updating and transformation) and connected
components analysis. Hence, the 3D grid representation satisfy the real-time demands of a multi-

~
Y[ ]

~
-4

2
-3

-~

3 4
-

x[m]

(c) Part of 3D image

Fig. 8. A experimental environment, the prefiltered
echoes and a part of the resulting 3D image
(range in meters).

400

P.G. Auran and O. Silven

beam sonar like the Simrad EM 1000 using lowcost hardware. It is a promising approach towards
autonomous underwater perception, and the research is continuing, with main focus on feature
extraction and object description from connected
3D echo clusters.
6. ACKNOWLEDGEMENTS
The authors wish to thank Donald P. Brutzmann, Naval Postgraduate School, USA and
Geir Mathisen, SINTEF Automatic Control, Norway, for their helpful support with information
and software regarding their research. This research was supported by NorFA (grant 94.30.150O) and the Research Council of Norway (grant
100626/410), greatly appreciated by the authors.
7. REFERENCES
Auran, P.G. and K. Malvig (1996a). 3D Sonar
Based Sensing for AUVs: Realtime Experiments within the Uncertainty Grid Framework.
Accepted for presentation at IFAC'96.
Auran, P.G. and K. Malvig (1996b). Realtime Extraction of Connected Components in 3D Sonar
Range Images. Submitted to CVPR'96 (IEEE
Computer Society Conference on Computer Vision ~ Pattern Recognition).
Bajcsy, R. (1988). Active perception. Proceedings
of the IEEE, 76 (8).
Balchen, J.G. and S. Yin (1994). The MOBATEL
Project - Model Based Teleoperation of an Underwater Vehicle Over a Narrow Band Communication Link. Technical report. The Norwegian
Institute of Technology.
Be~gem, O. (1993). Bathymetric Navigation of
Autonomous Underwater Vehicles using a
Multibeam Sonar and a Kalman Filter with
Relative Measurement Covariance Matrices.
Dr.scient thesis. University of Trondheim, Norway.
Blidberg, R., (ed.) (1987). Workshop Results
from the United States / France Cooperative
Program in Ocean Systems Technology: Autonomous Underwater Vehicles ~4 Underwater
Robotics.
Brutzmann, D.P. (1994). A Virtual World for an
Autonomous Underwater Vehicle. PhD thesis.
Naval Postgraduate School, USA.
Brutzmann,
D.P.,
M.A.
Compton
and Y. Kanayama (1992). Autonomous sonar
classification using expert systems. In: Proceedings of the IEEE Oceanic Engineering Society
OCEANS 92 Conference.
Durrant-Whyte, H.F. (1988). Sensor models
and multisensor integration. The International
Journal of Robotics Research, 7 (6).
Elfes, A. (1986). A Sonar-Based Mapping and
Navigation System. Autonomous Mobile Robots

Annual Report 1985, Mobile Robot Laboratory,

Carnegie-Mellon University.
Elfes, A. (1990). Occupancy Grids: A Stochastic
Spatial Representation for Active Robot Perception. In: Proceedings of the Sixth Conference
on Uncertainty in AL
Elfes, A. (1991). Dynamic Control of Robot Perception using Stochastic Spatial Models. In:
Proceedings of the International Workshop on
Information Processing in Mobile Robots.
Fossen, T.I. (1991). Nonlinear Modelling and Control of Underwater Vehicles. Dr.ing thesis. The
Norwegian Institute of Technology.
Leonard, J.J., H.F. Durrant-Whyte and I.J. Cox
(1992). Dynamic Map Building for An Autonomous Mobile Robot. The International
Journal of Robotics Research, 11 (4).
Maksym, J.N., A.J. Bonnet, C.A. Dent and G.L.
Hemphill (1983). Machine Analysis of Acoustical Signals. In: Proceedings of the NATO Advanced Research Workshop on Issues in Acoustic Signal/Image Processing and Recognition.
Mandelbaum, R. and M. Mintz (1994). Sonar Signal Processing using Tangent Clusters. In: Proceedings of the IEEE Oceanic Engineering Society OCEANS 94 Conference.
Mathisen, G. (1993). A System for Mapping of
a Partly Structured Environment for an AUV.
In: Proceedings of the 8th International Symposium on Unmanned Untethered Submersible
Technology.
Matthies, L. and A. Elfes (1988). Integration of
Sonar and Stereo Range Data using a GridBased Representation. In: Proceedings of the
IEEE International Conference on Robotics
and Automation.
Moran, B.A., J.J. Leonard and C. Chryssotomidis
(1993). Geometric Shape from Sonar Ranging.
In: Proceedings of the 8th International Symposium on Unmanned Untethered Submersible
Technology.
Moravec, H.P. (1986). Three Dimensional Images
from Cheap Sonar. Autonomous Mobile Robots
Annual Report 1985, Mobile Robot Laboratory,
Carnegie-Mellon University.
Ol'shevskii, V.V. (1978). Statistical Methods in
Sonar. Consultans Bureau, New York.
Simrad (1995). Simrad EM 1000 Hydrographic
Echo Sounder (report P24156E). Simrad Norge
A/S. Horten, Norway.
Smith, R., M. Self and P. Cheeseman (1987). A
Stochastic Map for Uncertain Spatial Relationships. In: Proceedings of the International Symposium on Robotics Research.
Urick, R.J. (1983). Principles of Underwater
Sound (3rd. ed.). McGraw-Hill.
WesMar (1970). Model SS150 Sonar Operations
Manual. Western Marine Electronics, Inc. Seattle, USA.