IEEE Robotics & Automation Magazine - March 2012 - 77

Navigation
The primary requirement of this observation program is
the ability to revisit benthic sites and image the same location on the seafloor. Revisiting the location of a single
image in surveys spaced out over a number of years is
likely to prove difficult even with high-end navigation
suites. The use of the dense grids allows an area to be revisited with a high degree of certainty as the majority of a
25 m 3 25 m patch of the seafloor is likely to overlap
between dives even if there is some offset in the estimated
vehicle location as might be expected when using a standard
GPS receiver. The broad survey grids, on the other hand, are
not designed to be revisited precisely but are meant to capture spatial variability within a particular dive site. A
standard set of oceanographic navigation instruments is,
therefore, sufficient for our purposes, although care must be
taken with calibration of the instruments and the manner in
which the navigation data are fused.
We operate an ocean-going AUV called Sirius capable
of undertaking the high-resolution, georeferenced survey
work [14]. This platform is a modified version of a midsize
robotic vehicle called SeaBED built at the Woods Hole
Oceanographic Institution [15]. This class of AUV has
been designed specifically for relatively low-speed, highresolution imaging and is passively stable in pitch and roll.
The submersible is equipped with a full suite of oceanographic sensors (see Table 1).
Real-Time Navigation
Our vehicle is equipped with a single-band GPS receiver,
a Doppler velocity log (DVL), a depth sensor, a magnetic
compass with integrated roll and pitch sensors, and an
ultrashort baseline (USBL) acoustic positioning system
deployed by the support vessel. The observations of
velocity provided by the DVL are combined with the
observations of attitude and depth using an extended
Kalman filter [14]. The USBL observations, consisting of
range and bearing measurements between the vessel and
the vehicle, are collected on the surface and are sent
together with the ship's position and attitude to the vehicle using the USBL's acoustic modem. These observations are received by the vehicle and fused into its
onboard navigation filter. The heading reference used is
sensitive to the magnetic signature of the rest of the vehicle [16], which can introduce distortions of several
degrees into the heading estimate. Even when soft and
hard iron calibrations are performed, persistent headingdependent errors of O (1°) are possible. While adequate
to perform linear transects or broader acoustic surveys
(particularly, when aided by acoustic positioning from a
long baseline or USBL), the magnitude of these errors
makes an intended dense mow the lawn pattern with
reciprocal, closely spaced, parallel track lines that are difficult for the vehicle to complete. We have recently
shown that it is possible to derive a heading-dependent
correction to the magnetic compass using visual data that

can enable a compass-equipped AUV to perform dense
visual coverage of a seafloor patch of approximately
50 m 3 75 m with 50 parallel track lines [17]. This has
resulted in a navigation suite that is capable of meeting
the requirements for repeated surveying of the permanent reference sites.
Simultaneous Localization and Mapping
To generate accurate models of the seafloor, it is important
that the estimated vehicle trajectory is self-consistent with
respect to the data being collected during each survey. We
employ visual SLAM to optimally fuse uncertain navigation estimates and visual observations [11]. This allows us
to further refine the estimated vehicle trajectory using the
environmental data, including high-resolution imagery
and multibeam sonar, collected during the survey. Cameras are capable of high-resolution observations so that
if the same scene is imaged from different positions, it is
possible to determine the relative poses of the cameras
using observations of features in the scene. These constraints are fused into the vehicle's navigation solution
to further refine the vehicle's estimated trajectory. Examples of loop closures identified in a dense survey are shown
in Figure 2(a).
To allow the survey data to be compared across years, it
is important that the annual surveys are coregistered. The
real-time navigation suite, including USBL observations, is
sufficient to position the vehicle within a meter of its
intended survey location, particularly in shallow water.
We have shown how loop closures can be identified in successive dives using standard SLAM techniques when the
time span between dives is short [10]. However, over the
course of a year or more, substantial changes in the benthos have often occurred and normal image features used
by our SLAM system do not reliably find matches. We are
currently working on developing multiresolution matching
techniques that use sonar data to provide gross registration
across years from major morphological features. Finerscale registration will need to account for variability in
the benthos itself and is an area of active research that
will exploit recent developments in the areas of image
change detection.
Delivering Data Products
We have demonstrated the ability of these AUV systems to
collect the type of data required to support the observation
program outlined herein. However, there is also a requirement to bridge the gap between the in situ observations
provided by the AUV and the information required to
answer specific scientific questions concerning changes in
marine habitats. The sheer volume of data available to
support these studies requires that much of the processing
and data integration be automated to avoid bottlenecks
associated with manual interpretation of imagery and associated data products. The data are currently available
online through the IMOS electronic marine infrastructure
MARCH 2012

IEEE ROBOTICS & AUTOMATION MAGAZINE

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - March 2012