Experiments with Driving Modes for Urban Robots

1

• Experiments with Driving Modes for Urban Robots
• Carnegie Mellon University
• TRM team
• JPL
• USC
• IS Robotics
• Oakridge Natl. Lab.

2
(No Transcript)
3
(No Transcript)
4
(No Transcript)
5
(No Transcript)
6
Algorithm Overview
To driving subsystem
Goal selection Initial template
Omni image
Virtual image
Predicted template location
Local search
Affine registration
Expanded search
Global search
7
Outline

• Experiments Description
• Performance Summary
• Practical Issues
• Template Selection
• Template Update
• View Transfer
• Recovery Modes
• Lighting Issues
• Failure Modes and Remaining Issues

8
Experiments Description

• Ft. Sam Houston experiments
• Prior testing at CMU

9
Examples

• Halfway toward the back of the building
• Turning toward the corner near the exterior stairs

10
Experiments Performance Summary

• Designation distance from 60m (back of building)
  to lt 10m (exterior staircase.)
• Maximum obstacle drop 10 curb, 10 obstacles
  (bricks, rocks, etc.)
• Terrain roughness Approx. 1 roughness
  amplitude.
• Slope variation Approx. 20o.
• Lighting conditions 8am to 8pm sun orientation.
  Bright sun to cloudy conditions.
• Speed from 0.8 m/s on straight and shallow arcs
  to 0.2m/s for tighter arcs. Turns at 30o/s.
• Image size Virtual image 300 by 200 pixels, 90o
  field of view.

11
Template Size
12
Template Size
50 pixels
40 pixels
30 pixels
13
Computation Time

• Computation time as function of template size
  (Pentium 166MHz vision stack.)
• Distribution
• Correlation 40
• Affine tracking 20
• Unwarping 10
• Smoothing, blooming removal, transformations
  30

14
Template Update

• Large variations in scale and orientation
  Periodic update of template
• Threshold for template re-initialization current
  size gt 1.5 initial size
• Heuristic for template update Shrink template
  toward the center.
• Servoed direction is maintained toward the center
  of the target.

15
Template Update Example
Initial target
Target reaches maximum change in scale in the
horizontal direction
Target is re-scaled horizontally
16
Template Update
Target adjustment
17
Template Update
18
View Transfer

• Virtual image /- 45o
• View transfer Virtual panning using omnicam
  unwarping
• Virtual views pre-computed at 45o and 20o
  intervals
• View transfer has near-zero cost

Target moved to virtual image 2, 45o off
Target
Target inside virtual image 1
1
2
Omnicam virtual spherical view
19
View Transfer Example

• Transfer example from 90o turn toward the corner
  of the building near the exterior stairs.

Target moves out of current virtual view
Target re-initialized in closest virtual view
(45o virtual camera pan)
20
Recovery Modes

• Full 6-dof tracking sensitive to rough motion
• Obstacles
• Fast point turns
• Direct correlation search added 32 pixels around
  predicted location of the template.
• Local search increased to 64 pixels around
  predicted location if target is lost
• Additional search in /- 45o region

21
Recovery Example
Normal tracking search window is /- 16 pixels
around current target position.
Recovery search search window is increased to
/- 32 pixels. Box shows region swept by the
upper left corner of target in the search region.
Recovered target after template initialization.
22
Lighting Issues

• Harsh lighting conditions Blooming
• Solution Interpolation algorithm
• Evaluation Tracking in a variety of weather
  conditions

Interpolation algorithm
Blooming artifact
23
Issues

• Integration with other behaviors
• Computational issues
• Frequency of global search events
• Lighting issues
• Dynamic mixing conventional cameras/omnidirectiona
  l cameras
• Use of range info from stereo
• Use of robot state info