SLAM Summer School 2006Dr. Stefan Williams1 CAS Marine Systems Stefan Williams, Oscar Pizarro, Ian Mahon, Paul Rigby, Matthew Johnson-Roberson ARC Centre.

Dr. Stefan Williams1 SLAM Summer School 2006

CAS Marine SystemsStefan Williams, Oscar Pizarro, Ian

Mahon, Paul Rigby, Matthew Johnson-Roberson

ARC Centre of Excellence for Autonomous Systems

School of Aerospace, Mechanical and Mechatronic Engineering

The University of Sydney


Overview

• Unmanned Underwater Vehicles

• Reef Monitoring

• Navigation Sensors

• Terrain Aided Navigation

• SLAM

• Classifications

• Future Challenges

• Conclusions


Unmanned Underwater Vehicles (UUVs)

• Constraints• No GPS• Low cost IMU• Unstructured

Terrain

• Research Challenges• Sensing and

Perception• Localisation and

Mapping• Control

Odyssey AUV (Image courtesy of MIT)

ROV Tiburon (Image courtesy of MBARI)

Phoenician wreck mosaick (Image courtesy of WHOI)


Reef Monitoring

• This project is developing algorithms and methods for modelling marine environments

• Long term goals are to provide a set of tools to support research into the ecologically sustainable use and protection of tropical coral reefs and other marine habitats


Reef Monitoring

• Recently returned from deployment at Capricorn Bunker Reef Group in the Southern GBR

• Deployed AUV at numerous locations to evaluate performance of system and collect data sets

• 21 data sets collected undertaking a variety of mission profiles

One Tree Island

Fitzroy Reef


Biodiversity Assessment

• AIMS are currently conducting biodiversity assessment in collaboration with the Western Australian Museum off the North West Cape

• Techniques include dredge sleds, drop cameras and towed video sleds

• Particular focus on sponge habitats beyond diver depths

• Assessment for designation of Marine Sanctuaries


Surveys off Ningaloo Reef

• Invited by AIMS to undertake surveys off the coast near Ningaloo Reef

• Surveys of back reef and sponge beds in up to 80m depth targeting areas based on Mandu Roughness Values shown

Ningaloo Reef


The AUV platform

• Fully autonomous operation, no tether

• Control of vehicle performed using on-board computer

• Sensors include • Sonar (imaging and

fwd obstacle avoidance)

• Vision (stereo)• DVL• Compass• Pressure

• Mission Time ~4 hours (up to 8 hours with current housing)


Vehicle Specifications

Vehicle Specifications

Depth rating 700m

Size 2.0 m(L) x 1.5 m(H) x 1.5 m(W)

Mass 200kg

Maximum Speed 1.2 m/s

Batteries 800 Wh Li-ion pack

Propulsion 3x150 W brushless DC thrusters

Navigation

Attitude/Heading Tilt (±0.5o), Compass (±2 o)

Depth Paroscientific pressure sensor, (0.01 %)

Velocity RDI Navigator ADCP (1-2mm/s)

Altitude RDI Navigator

USBL TrackLink 1500 HA (0.2m range, 0.25 o)

Optical Sensing

Camera Proslica 12bit 1360x1024 CCD

Lighting 2 x 200 Ws strobe

Separation ~1 m between camera and light

Acoustic Sensing

Imaging sonar Tritech Seaking

Obstacle Avoidance lmagenex 852 Echo Sounder

Other Sensors

CT Seabird 37SBI


AUV Navigation

• A common navigation sensor used in most commercial AUV’s and many ROV’s is the Doppler Velocity Log (DVL)

• A DVL transmits 3 or 4 sonar beams downward from the AUV to measure the speed along and across the AUV track

• The AUV controller integrates heading, attitude and velocity into a dead-reckoned position estimate


AUV Navigation

• Both GPS and acoustic positioning systems can be used to assist in determining AUV position. There are several methods for this including:

• Long Baseline Systems (LBL)

• Short baseline systems (SBL)

• Ultra short baseline system (USBL)


AUV Navigation

• The ultrashort baseline system (USBL) requires only one transducer to be installed.

• This makes the ultrashort baseline system more portable


DVL and USBL Performance

• Data from a recent mission illustrates the overlapping grid patterns used for dense habitat modelling


DVL and USBL Performance

• Calibration of DVL and compass results in significantly improved error characteristics (on the order of 2% of distance travelled)

• Also working on methods for automatic calibration of vehicle and tracking sensors


Terrain Aided Navigation

• Terrain elevation maps available for some deployment areas

• Observations of altitude can be used to bound likely position of the vehicle


Laser Airborne Depth Sounder (LADS)

• Can be deployed in suitable conditions where water depth is 2-50m

• The infra-red component reflects from the surface

• The green component penetrates the water and reflects from the sea floor

• Depth determined from difference between the two beams


Laser Airborne Depth Sounder (LADS)

• LADS very suitable suited for reef environments

• We have data sets available for GBR

• Can be as accurate as +/-5m horizontally +/- 0.5m depth


Sydney Harbour Demonstrations

• Sydney Harbour presents an ideal environment in which to validate these algorithms

• Detailed bathymetric maps of the harbour are available


Sydney Harbour Bathymetry

• Sydney Harbour Bathymetry from DSTO Shallow Water Survey

• Bathymetric data collected using multi beam echo sounder

• Resolution of 1m over extent of inner harbour

Darling Harbour

Harbour Tunnel


Ship transect

• Ship data from DSTO, including GPS position and depth soundings taken at 5s intervals, during transect of the Harbour

• Particle based localisation and tracking of ship using depth soundings has been demonstrated using logged data


Particle Tracking of Ship Transect


Particle Tracking of Ship Transect Lost


The SLAM Problem

EstimatedLandmark

TrueLandmark

True Vehicle Path

EstimatedVehicle Path

CorrelatedLandmark Errors

Simultaneous Localisation and Map Building (SLAM)

Start at an unknown location with no a priori knowledge of landmark locations

From relative observations of landmarks, compute estimate of vehicle location and estimate of landmark locations

While continuing in motion, build complete map of landmarks and use these to provide continuous estimates of vehicle location


The Estimation Process

Recursive three stage update procedure using Extended Kalman Filter (EKF)

PredictionPrediction

– Use vehicle model to Use vehicle model to predict vehicle positionpredict vehicle position

ObservationObservation

–Take feature observation(s)Take feature observation(s)

UpdateUpdate

–Validated observations Validated observations used to generate optimal used to generate optimal estimateestimate

–Initialise new targetInitialise new target

F0

0xv

Vehicle Path

0x2

0x3

0x1

0x4

F0

0xv

0x2

0x3

0x1

0x4

Vehicle Path

F0

0xv

0x2

0x3

0x1

0x4

Z1

Z2

Z3

Z4

Vehicle Path

F0

0xv

0x2

0x3

0x1

0x4

0x6

0x5

Vehicle Path


SLAM

1. Initialize vehicle at origin

2. Take sonar observation of Range Bearing

3. Initialize estimate of features and project into visual frame

4. Identify visual feature within sonar footprint

5. Track visual features using KLT and provide elevation azimuth observations to filter


Fusing Vision and Sonar

• Using the sonar returns to initialise visual features

• Features are then tracked from frame to frame using KLT

• Observations provided to SLAM algorithm to build terrain model and estimate vehicle motion

• KLT features are not stable over long term


Terrain Model

• Sonar returns outside of visual frames are used to generate a coarse surface map

• Visual frames are projected onto surface

• Correspondence at seams suggests that algorithm is performing well

• Resulting terrain model


Terrain Models


Stereo Imaging

• High resolution stereo pairs allow local surface to be reconstructed

• Alternatively, features in images can be identified and position relative to vehicle computed

• Developing techniques for integration into SLAM framework


Stereo Imaging


Stereo Imaging


Stereo Imaging


Sponge Beds at 80m off Ningaloo


Visual Mosaics at 40m off Ningaloo


Terrain Classification

• Techniques for automatic terrain classification

• Combined with terrain models provide a mechanism for estimating cover

• Preliminary work looked at the use of Gray Scale Histograms and mean RGB colour



• Selection of a richer description of texture is required

• Gabor filter is used a varying scales and orientation to identify patches of coral, sand and other terrain in images

• Investigating combination of visual and sonar data for classification




• Autonomous Systems Roles:• Wide-Area

Surveillance• Situational

Awareness• Air/Land/Sea

Operations • Tactical Strike• Mine Hunting• Littoral Zone Support

and Sensor Payload Delivery

• Research Challenges:• Multi-Sensor Data

Fusion• Multiple Platform

Control• Autonomous

Operations• Systems of Systems

Future Challenges: Defence Applications


Future Challenges: Integrated Ocean Observatories

• Autonomous Systems Roles:• Wide-Area Surveillance• Situational Awareness• Air/Land/Sea Operations

• Research Challenges:• Multi-Sensor Data

Fusion• Modelling Ocean

Processes• Multiple Platform

Control• Autonomous Operations• Systems of Systems


• Autonomous Systems Roles:• Exploration• Mapping• Extraction

• Research Challenges:• Multi-Sensor

Data Fusion• Adaptive

Sensing

Future Challenges: Resource Exploration


Conclusions

• Terrain aided navigation is feasible• Extension to unstructured terrain

map building on Autonomous Underwater Vehicle (AUV)

• Deployment in various fielded situations

• Applications• Ecological Monitoring• Mineral Exploration• Mine Hunting• Littoral Zone Support and Sensor

Payload Delivery


Acknowledgements

• This work is supported in part by the Australian Research Council (ARC) and the New South Wales Government

• Additional funding has been received from our industry partners, including• BAe Systems ATC• Sonartech Atlas• The Great Barrier Reef Research

Foundation• The Commonwealth of Australia• DSTO


SLAM Summer School 2006Dr. Stefan Williams1 CAS Marine Systems Stefan Williams, Oscar Pizarro, Ian Mahon, Paul Rigby, Matthew Johnson-Roberson ARC Centre.

Documents

slam summer school

ningaloo reef slide

stefan williams4 reef

stefan williams3

stefan williams8

sbi slide

ningaloo reef surveys

stefan williams7 surveys