Terrain Modelling for the Mars Exploration Rovers

1
Terrain Modellingfor theMars Exploration Rovers

• John R. Wright, Brian Cooper, Scott Maxwell,
  Frank Hartman, Jeng Yen, and Jack Morrison
• Jet Propulsion Laboratory
• California Institute of Technology
• Under Contract to NASA

2
Why Do Terrain Modelling

• Understanding the in-situ conditions of the rover
  is necessary for planning optimal use of
  scientific instruments while minimizing risk
• Three-dimensional visualization offers optimal
  understanding

3
Goals

• Process data from multiple sensors, platforms,
  and missions
• Coregister the models
• Merge the models
• Extract models formatted for various applications
  (meshes, height maps, etc.)
• Maintain maximum fidelity and accuracy
• Support terrain exploration and interaction
• Results have been mixed

4
The MER Rovers and Imagers
NavCams
PanCams
Front HazCams
Rear HazCams
5
Operational Paradigm
Site 3
Site 2
Landing Site
Site 4
6
Original Stereo Images from NavCam
Right
Left
7
Wedge Generation

• Each pair of stereo images from each imager is
  processed to generate a wedge of terrain model
• Stereo Correlation is performed to generate a
  disparity map
• Disparity and camera model information is used to
  generate a range map
• Camera pointing information is then used to
  convert range to an (X,Y,Z) triplet for each
  pixel
• The cloud of points thus generated forms the
  basic model unit used to build complex models of
  the entire area of exploration

8
Registration

• Registration of wedges to each other or to a base
  model is required to correct pointing and
  localization errors
• Iterative Closest Points algorithm seems ideal
  for this application
• Should be able to handle different resolutions
• Difficult to verify with real data
• Likely that errors in localization cause subtle
  problems with shape that lead to poor
  registration
• Currently we use camera pointing for registration
  purposes
• Localization remains an issue

9
Merging

• Models/wedges are merged in an octree
• Samples are assumed to have a volume equivalent
  to resolution
• Samples are assumed to be points when generating
  an output model
• A forest is used to support merging of multiple
  octrees in a structure that supports modifying
  registration transforms after the fact
• Works great

10
Mesh Generation

• The merged cloud of points is processed to
  generate a triangle mesh representing the terrain
• First used Marching Triangles algorithm to
  utilize the actual point set
• Dependent on excellent registration which is
  problematic
• Currently we mesh in image space
• LODs generated from subsampling in image space
• Tiling is supported by the octree and provides
  for better rendering performance

11
The Product - Spirit Landing Site
12
The Product - Spirit Landing Site
13
Wireframes
14
Rover-Eye View
15
Height Maps

• Quadtrees extracted directly from the original
  octree
• Multi-resolution supported nicely
• Generates a two-banded height image
• First band is uninterpolated
• Second band is interpolated to fill holes
• Works great
• Height maps are used for rover settling and for
  instrument/terrain collision checking

16
Associated Height Map
17
HazCam Models
18
HazCam Models
19
Open Issues and Acknowledgements

• Need better methods to extract meshes from the
  merged models in the octree, giving priority to
  high-resolution data with minimal or no
  smoothing
• Need better registration methods that are more
  tolerant to shape variation
• Need modelling methods that are more accurate
  with long baselines
• All processes must be very autonomous for rapid
  turnaround with no hand tweaking

The work described herein was performed at the
Jet Propulsion Laboratory, California Institute
of Technology, under contract to NASA.