42 MAY 2010 INTERNATIONAL WATER POWER & DAM CONSTRUCTION
UNDERWATER INSPECTION
INTELLIGENT CONTROL ARCHITECTURE
Ictineu AUV is endowed with intelligent software responsible for the
navigation, guidance and control of the vehicle. The mission is pro-
grammed using mission control language (MCL), an imperative lan-
guage which compiles a Petri-net based representation of the mission
[5], useful for pre-mission verication and real-time mission execu-
tion. During the run time, the mission controller is in charge of the
sequential and/or parallel execution of tasks.
The system includes tasks for checking safety alarms, logging
sensor data, waypoint guidance, path following, sensors enabling/
disabling, keeping a relative position with respect to the wall [6],
altitude control, etc. The result of the mission execution is a set of
log les containing all the data gathered by the sensors, including the
imagery, conveniently synchronised. This data is then post-processed
to build the maps.
GEO-REFERENCING THE ROBOT POSITION
During a mission, the vehicle estimates the trajectory by means of a
stochastic sensor fusion algorithm known as the extended Kalman
lter [7]. The lter estimates the vehicle state (position, heading and
velocity) and its corresponding uncertainty following a two-step
recursive process. First, a model is used to predict the vehicle motion.
Then, this predicted state is corrected by introducing new information
provided by the sensors.
In our particular application, the velocity measurements from the
DVL, the depth calculated from the pressure sensor and the heading
provided by the bre optic gyro are used to perform the state update.
As a result of this process, a dead-reckoning estimation of the vehicle
motion is obtained. The USBL system provides absolute measure-
ments required to geo-reference the vehicle as well as to bind the error
growth in the estimated position.
To operate, the USBL system makes use of a second Kalman lter
that constantly estimates the position and altitude of the transducer
through the measurements provided by the MRU and the DGPS. When
a USBL measurement is obtained (vehicle position referenced to the
sensor frame) it is composed with the current lter estimate (absolute
position and altitude of the sensor) to produce the vehicle position ref-
erenced to the world frame. This global position is then fused ofine
with the dead-reckoning position estimate, to compute the geo-refer-
enced robot trajectory to assist the mosaic building process.
PHOTO-MOSAI CS
Photo-mosaics are built ofine by stitching together all the images
gathered during the survey and making use of the navigation data
(robot pose). The process is done automatically using software devel-
oped by our team. First of all the algorithm processes each pair of
sequential overlapping images common features. This allows compu-
tation (up to scale) of the camera motion allowing alignment of the
images forming a global visual map (of the photo-mosaic).
It is worth noting that because this is an iterative process, the small
registration errors are accumulative, provoking a drift. Whenever the
robot re-visits an already mapped area (crossover), a loop is gener-
ated. This allows a non-consecutive image registration to be used
for globally aligning the photo-mosaic through a technique known
as bundle adjustment. Finally, the mosaic alignment is improved
through several iterations of crossover detection and optimisation.
The result is a globally aligned photo-mosaic.
EXPERIMENTAL RESULTS
In July 2008 several experiments were carried out upstream of the
Mequinenza dam on the Ebro river in Spain. This was part of a collab-
oration project with Ecohydros, a Spanish company that offers aquatic
ecosystems consulting services. The purpose of the experiments was to
produce several geo-referenced photo-mosaics of the bottom. This was
to provide visual validation for a sonar-based system being developed
by the company to detect zebra mussel colonies.
During the test trials, the Ictineu vehicle together with the USBL
system was mounted outboard of a boat and performed small sur-
veys in some areas of interest. The water turbidity forced navigation
close to the bottom to ensure sufcient image quality. This made it
difcult to control the vehicle and reduced the performance of the
DVL measurements (our device is unable to provide reliable velocity
estimates under 1m altitude). Moreover, the harsh environment, with
big slopes and large rocks, made the autonomous operation of the
vehicle impossible. For this reason, the Ictineu was set to ROV mode
and operated from the boat with the help of the real time feedback
from the imaging sonar, the vehicle cameras and the position estimate
from the USBL.
The results of the different dives were irregular. The unreliable DVL
data and the multipath affecting the USBL measurements made it dif-
cult to obtain a good position estimate. The quality of the captured
images was also irregular. It was necessary to perform manoeuvres
to avoid hitting large rocks present on the scenario and as a conse-
quence, the camera suffered abrupt changes in the altitude, sometimes
losing sight of the bottom because of the turbid waters. However,
despite all the commented issues, it was still possible to generate some
photo-mosaics of sufcient quality.
EXPERIMENTS AT THE PASTERAL DAM
The purpose of this test was to demonstrate the capacity of the system to
execute an inspection of a dam wall to search for cracks or other dam-
ages on the concrete. Nowadays, professional divers perform this task,
but they offer only a limited view of the submerged structure, which
makes it difcult to clearly determine the position of the defects.
The proposed method overcomes this problem by generating a geo-
referenced mosaic from which a technician can easily determine the
exact location of the spot of interest. Moreover, the fact that the under-
water vehicle can execute this task autonomously makes it possible to
increase the frequency of the inspections.
During the tests carried out in February 2009 at the Pasteral Dam
(Ter river, Spain) the Ictineu executed autonomously a survey trajec-
tory covering a rectangular area of approximately 4x10m with the
forward looking camera. The USBL system was placed on a buoy and
Figure 4 – One of the mosaics produced during the Mequinenza trials.
A mussel colony can be observed at the center of the image
Figure 5 – Mosaic of a wall in the Pasteral hydroelectric dam