osgAR: a Scene Graph with Uncertain Transformations

1
osgAR a Scene Graph with Uncertain
Transformations

• Enylton Machado CoelhoBlair MacIntyre
• Augmented Environments Lab, GVU - CoC
• Simon Julier
• Naval Research Lab

2
Topics

• What is AR?
• Registration Error
• Scene graphs osgAR
• Components
• Limitations
• Current future work

3
Augmented Reality

• Augment, not replace, the physical world with
  computer-generated objects

4
AR in Maintenance

• Microvision Honda trial
• Access to maintenance library

Reference
www.microvision.com/hondatrial
5
AR Using Visually Coupled Head-worn Displays

• Combine graphics with physical world

6
Registration Error

• Misalignment between the computer generated
  graphics and the physical object

7
Registration Error

• Commonly used approach
• Better trackers
• More accurate modeling and calibration
• Faster computers
• Not practical in real situations
• Trackers may break
• Knowledge will never be complete

8
Registration Error

• Our approach
• Assume errors will always exist
• Estimate resulting registration errors
• Use error estimates to drive the graphics
• Developers concentrate on the intent of the
  augmentations
• Decouple from tracker characteristics

9
Registration Error

• Changing what is being displayed ameliorates the
  registration error

LABELS
10
Registration Error

• Once the registration error can be estimated,
  different augmentation techniques can be tested
• Estimating the error at run time is the hard part

Reference
www.microvision.com/hondatrial
11
Topics

• What is AR?
• Registration Error
• Scene graphs osgAR
• Components
• Limitations
• Current future work

12
Scene Graphs

• Rigid transformations
• Hierarchical representation
• Widely adopted
• Inventor, Java3D,

13
Scene Graphs with Uncertainty

• Error estimates are propagated down the graph

14
Previous WorkStatistical Error Estimation

• Individual vertices
• 2D screen region

Reference
VR02 Estimating and Adapting to Registration
Errors in AR Systems
15
osgARArchitecture

• Based on OpenSceneGraph (www.openscenegraph.org)
• Extended to Augmented Reality
• Support for AR
• Uncertainty

Reference
ISMAR04 osgAR A Scene Graph with Uncertain
Transformations
16
osgARAR Support

• Video in the background
• Tracker support
• VRPN
• ARToolkit
• 2D interface manager

17
osgARComputing the Estimate

• Model the Uncertainty as a Gaussian
• Adds a covariance matrix to the original 4x4
  matrix transformation

18
Bounding Regions

• Inner Always inside the object
• Outer Contains the object

BOUNDER
19
osgARExposing the Estimate

• Region polygonal representation of the regions
• Assessment single value corresponding to the
  objects registration error

20
osgARExamples of Using the Estimates

• Region
• Label Placer
• Bounder
• Assessment
• LOE

CALLOUTS
Reference
ISAR00 Adapting to Registration Errors Using
Level of Error (LOE) Filtering
21
Multiple TrackersTransformation Combiner

• Multiple paths to a transform
• Callback function picks which to use
• Parameter list of error estimates
• Return which path and estimate to use

22
Multiple TrackersTransformation Combiner
COMBINER
Base
Sensor
Camera
COMBBOUNDER
23
osgARArchitecture

• AR Support
• Estimate
• Computation
• Expose
• Examples
• Multiple Trackers

24
Observations

• Should use shortest path in graph
• Camera tracker
• Hack reset error at camera
• Head/object tracked with same sensor
• Solution more elaborate bookkeeping/traversal
• Leverage redundant information

25
Camera Uncertainty
attached to the world
attached to the camera
26
Pending Transforms

• Transformations other then tracker
  transformations are updated by the system

PENDING
27
Current and Future Work

• Generic model that computes the optimal
  registration error estimate
• Exploit the redundancy in the system
• Possibility of adding interaction
• Applicability and limitations of current computer
  graphics models

28
Acknowledgements

• Members of the AEL and GVU for many discussions
  and ideas
• ONR grant N000140010361

FOR MORE INFO...
www.cc.gatech.edu/ael
29
Error Estimation

• Compute statistical properties for each vertex of
  an object
• Aggregate these estimates per object

30
Statistical Error Estimation(Simon Julier, NRL)

• Unscented Transformation
• Easy to implement
• More accurate than linearization
• Fast

31
Error Estimate Aggregation

• 2D Convex Hull
• Project error bounds on 2D screen
• Compute convex hull

32
osgAR Traversals

• Optimizer
• 3D uncertainty propagation
• Registration error computation