stanford university, nasa ames, mbari range data ... · stephen rock at [email protected], or visit...
TRANSCRIPT
[1] Hammond, M., Clark, A., et. al., “Automated Point Cloud Correspondence Detection for Underwater
Mapping Using AUVs,” OCEANS, Washington DC, Oct. 19-22, 2015.
[2] Images taken from: http://stocktouch.com/wp-content/uploads/2012/02/iceberg-poster.jpg (Iceberg),
www.esa.int (Comet), and www.CartoonStock.com, ID tbrn213 (Cartoon Canyon Wall)
This work was funded under NASA ASTEP Grant NNX11AR62G. The authors would
also like to thank the Monterey Bay Aquarium Research Institute for the data, ship
time, and technical support. Particular thanks go to Brett Hobson, Hans Thomas,
Rob McEwen, and Rich Henthorn.
Create accurate 3D reconstructions of natural terrain, subject to
unknown terrain motion and/or substantial vehicle inertial
estimation errors.
Mapping Asteroids and Icebergs:
Range Data Correspondence Detection for Natural Terrain
Objective
Matches were determined by the quantity of RANSAC inliers. Varying
the inlier threshold trades match quantity versus match accuracy. A
Hough transform successfully identifies false matches.
Results
For decades, spacecraft have been utilizing star trackers for attitude
determination. Their constellation matching algorithms have been
designed for low computation and low memory requirements.
How Star Trackers Work (& How it Applies to Range Data):
1. Take an image of stars (collect range data)
2. Locate stars within the image (locate point cloud features)
3. Search for a matching constellation in a starfield database (identify
matching constellations of features in the “database” dataset)
Constellation Matching
BibliographyAcknowledgments
For further information, contact Ashley Clark at [email protected],
Stephen Rock at [email protected], or visit arl.stanford.edu
Further Information
Initial Experiments Ongoing Improvements
Ashley Clark and Stephen RockStanford University, NASA Ames, MBARI
Sonar data was collected by an
autonomous underwater vehicle, along
two passes around Soquel Canyon in
Monterey Bay, CA.
Feature Detector: SIFT
Feature Descriptor: SIFT
Feature Match Algorithm: RANSAC
Outlier Rejection: Hough Transform
Problem Setup
Ground truth is shown in a green band. Accepted matches
are shown in blue. Rejected matches are shown in red.
Problem Description
Motivation: Asteroid exploration can provide clues about the origin of our
solar system, but the uncertainty of asteroid spin rates and spin axes makes
3D reconstruction and subsequent terrain-relative navigation challenging. This
research develops algorithms for asteroid missions and tests them on
comparable Earth-based scenarios, such as drifting iceberg exploration and
underwater vehicles with poor dead reckoning.
Issues:
• Poor Terrain-Relative Inertials
− No GPS
− Drifting Target
− Poor Sensors
• Lack of Distinctive Features
− Natural Terrain
− No Man-Made
Navigation Aids
Where does the problem come from?
ICP is commonly used to align point clouds, but
substantial inertial errors can cause ICP to
converge to local minima.
People have initialized ICP before. Why not
use their methods? Most ICP initialization
routines rely on easily identifiable and well
localized features, i.e. from manmade objects
with sharp corners or pre-placed fiducial markers.
Will the algorithm work on noisy
range data from a real system?
Success!
Method Overview
High Level Algorithm:
Use GraphSLAM to correct
the warped point cloud.
New Correspondence
Detection Algorithm1:
To detect correspondences between one set of range data and another, use image
processing to find matching subclouds.
Warped
Point Cloud
Detect
Correspondences
Run ICP on Each
Correspondence
GraphSLAM
Using ICP
Offsets
Segment Data
into Subclouds
Detect
Features
Describe
Features
Identify Feature
Matches
Quantify Subcloud
Match Quality
14
Outlier
Rejection
For Each Subcloud For Each Pair of Subclouds
Initial experiments demonstrate the success of the algorithm in real
terrain, but the reliance on RANSAC for matching presents three issues:
• Natural terrain can have relatively few distinctive features; RANSAC
depends on a large quantity for successful matching
• A small number of closely-matched features can indicate
correspondence; quantity of matched features is not a sufficient metric
• Computer memory is limited on spacecraft; reducing the minimum
necessary quantity of features is preferred
Challenges
Using constellation matching algorithms in lieu of RANSAC can reduce the
number of features needed per image by an order of magnitude
Benefits:
• Fast database lookup
• Order of magnitude
reduction of features
• Probabilistic model of
match quality