)N
ration,
| well-
'estrial
e-wise
le, our
| intra-
DEMs
ks) are
tified.
«s. For
wltiple
> robot
jixel in
for the
ing and
Camera
era bar
meters.
sists of
'etween
;ercent.
ereo) is
ent for
ods can
:clusion
ing-cost
2002).
for the
rio. The
s (Ohta,
rection,
raints in
he first
is very
thods of
Jensfelt,
indoor
ructured
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
environments as well (Krotkov, 1994; Sutherland, 1994: Olson,
2000; Cozman, 2000). Real-time applications are preferred
(Atiya, 1993). To assistant robot localization, landmarks are
selected or maps are built in the following applications:
Shimshoni, 2002; Mouaddib, 2002; Betke, 1997; Davison, 2002;
Olson, 2002.
1.3 Approach
Our goal is to generate terrain maps and orthophotos using
Navcam and Pancam panoramic stereo images to support
traverse design and to localize each rover by adjustment with
cross-site tie points.
The core of map generation is registration between intra-stereo
and inter-stereo. images and spatial interpolation. For an
unstructured extraterrestrial environment, features like edges
and surfaces rarely exist, thus we select interest points as our
features for matching.
Interest points between intra-stereo image pairs are matched
locally and verified globally. The verification of matching is a
global matching process of two steps: first, elimination of large
parallax outliers using a median filter in the vertical profile
(perpendicular to the scanline) by assuming piecewise
continuity, which is true for a natural terrain; second, detection
of small parallax outliers by triangulating all points in the X-Y
plane, back-projecting them onto the photo plane, and then
checking disordering nodes.
Interest points between inter-stereo image pairs are actually
matched in 3-D. For each point there are four observations; this
redundancy can be used to reliably eliminate outliers.
Instead of finding parallax for every point in the image plane,
which is inaccurate and unreliable for featureless areas, we
interpolate the terrain surface in 3-D using highly reliable
points. Kriging, for the close range, and Triangular Irregular
Network (TIN), for the far range, are used for spatial
interpolation.
Landmarks, such as rocks, are detected by projecting the
interpolated DEM back onto a number of corresponding images
and comparing the parallax difference. Rocks from different
sites are matched by considering measurement and localization
uncertainties. These rocks are then used as cross-site tic points
to adjust the rover location through rigid transformation and
bundle adjustment.
2. MAPPING WITH DESCENT IMAGERY (DIMES)
At each of the rover landing sites, Gusev Crater and Meridiani
Planum, three descent images (DIMES) were taken (from
around 1400m, 1100m, and 800m elevation), which were used
to form a vertical baseline configuration. Image parameters
were: size 1024x512, resolution around 1m (lowest image), and
coverage area lkmxlkm. Highly visible landmarks (15 for
Gusev and 19 for Meridiani) were manually selected as control
points in order to link the DIMES images to the MOC-NA (for
the X-Y coordinates) and the MOLA image (for elevation).
Then a bundle adjustment was performed to infer the
parameters of the DIMES images. These control points also
define a dual-directional bilinear transformation between the
lowest, middle, and highest DIMES images. These images are
then aligned by transformation, resampled to the same
resolution, and registered along the epipolar line.
PER EA
Figure 1. DIMES images from the Gusev site; DEM: and the
corresponding chromadepth map
The 3-D coordinates of the matched points are calculated via
spatial intersection. A small percent of the points are treated as
blunders and eliminated using correlation coefficients and local
terrain variations. The final DEM represents the general terrain,
as shown in Figure 1.
3. MAPPING WITH ROBOTIC IMAGERY
The mapping with robotic imagery involves the registration and
verification of intra-stereo and inter-stereo imagery as well
spatial interpolation with 3-D interest points. Figure 2 shows
typical Navcam images from Mars (inter-stereo images are
separated with black lines). Fórstner interest points (Fórstner,
1986) are extracted from these images as features. Their
number ranges from 300-1500 per image.
Figure 2. Overlap of typical Mars Navcam images
3.1 Intra-stereo Registration and Verification
Intra-stereo points are interest points linking intra-stereo images.
They are matched using block-matching and least-squares
matching (Wang, 1990) applied with constraints such as
epipolar and bi-directional uniqueness. The precision of
matched parallax can reach a 1/3 pixel level. The left-hand
image in Figure 3 shows an initial matching result.
Since the number of interest points per image is around the
number of image pixels per line n, and because for each interest
point only several other points along the epipolar line needs to
be checked, the overall matching process is O(»),, which can be
implemented in real-time with low cost.
To verify the match, parallaxes of matched interest points are
ordered in the row direction, as shown in Figure 4 (left). The
existence of outliers is obvious. Since an unstructured natural
terrain is generally piece-wise continuous, parallax is
monotonically decreasing from top to bottom. The distribution
of parallax can be represented with several pieces of curves that
can be derived by filtering the parallaxes with a median filter
and then approximating it with cubic b-splines. By thresholding
the parallaxes between matched pairs and the parallax curve,
extreme outliers can be eliminated. The threshold is a function
of distance and is set large enough to allow for the roughness of
1043
