The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
image pyramid was constructed consisting of five levels.
Starting with the original image, each subsequent level was
created by sub-sampling the previous level’s image smoothed
by a Gaussian filter. Interest points were generated by Fòrstner
operator (Fòrstner, 1986) at every image scale.
Figure 3. Workflow of the hierarchical stereo matching process
Matching started with the images of lowest resolution; results
were then transferred to the next higher level, with more interest
points being extracted and matched. The images, interest points
and orientation parameters were used as input for the stereo
matching process. At the lowest level, geographic locations of
interest points were estimated by assuming a flat terrain. This
enabled automatic pairing of interest points in stereo images.
The search radius was confined to the neighborhood of the
corresponding interest points, and matched points were selected
based on correlation coefficient values.
Automatic error detection is performed at each level by
eliminating outliers based on elevation distribution of
neighboring points. For each point, a small local DEM surface
was constructed from the matched points and modeled as a flat
terrain (may be improved to a plane). Then the standard
deviation of the plane estimation, a, was calculated. If the
residual of a point exceeded 2a, it was regarded as an error and
eliminated.
At subsequent levels, points from the previous level were
matched again to achieve higher matching precision. A TIN
(Triangulated Irregular Network) surface of parallax differences
was generated from these matched points using the Delaunay
triangulation method. This TIN was used to estimate the
corresponding tie points. To improve matching performance for
points located around the boundary of each CCD, HiRISE
imaging geometry was fully utilized. Instead of mosaicking
images from separate CCDs, we loosely stitched together the
TIN surface based on the best-fitting alignment derived from
interest point matching between adjacent CCDs.
After matching the interest points generated from the highest-
resolution images, 10-pixel grid points were defined to form a
basis for further matching. To generate a 1-m-resolution DEM
of the terrain, 3-pixel grid points were matched. For sub-meter
level DEM, dense matching was performed for every pixel in
the images of highest-resolution. Evenly distributed tie points
between the stereo images were selected from the set of
matched interest points to be used in the subsequent bundle
adjustment. The final DEM was generated after bundle
adjustment and elimination of matching errors.
3.2 Matching performance evaluation
We tested this process using a stereo pair of HiRISE images that
cover the Columbia Hills area of the Spirit rover landing site
(TRAJ301513 1655 and TRA_001777_1650). The
TRA001513 image was obtained on November 22, 2006. It is
centered at 14.6 °S latitude, 175.5°E longitude. It has 27.1
cm/pixel resolution and 80,000 rows. The TRA 001777 image
was taken on December 12, 2006. It has a resolution of 26.3
cm/pixel and 40,000 rows. Its extent is entirely covered by
TRA001513. The two images have a convergence angle of
19.8 degrees.
Level
Image
Scale
Point
Type
Residuals (pixel)
Mean
Standard
Deviation
Maximum
1
1/16
Interest
0.26
0.55
1.41
2
1/8
Interest
0.19
0.53
2
3
1/4
Interest
0.13
0.33
1
4
1/2
Interest
0.18
0.47
1.41
5
1
Interest
0.19
0.39
1
6
1
10-pixel grid
0.06
0.24
1
Table 1. Matching residuals at intermediate levels
A quantitative evaluation of matched points was conducted for
the hierarchical matching results. At each intermediate level, 16
points were randomly selected throughout the entire study area.
The automatically generated matching results were compared
with manually matched points. Table 1 shows the results of this
evaluation of matching accuracy. The highest mean residual
was found at the first level. However, this mean was still less
than 1 pixel, with maximum residual being 1.41 pixels. The
largest error, 2 pixels, was found at the second level. Since
interest points for levels 1 through 4 were projected onto higher
resolution images and adjusted by re-matching, the errors from
the previous levels were not propagated into subsequent levels.
While the mean residuals did not necessarily decrease over the
hierarchical process, they did remain at a reasonably low level
providing accurate-enough estimates of parallax differences for
use at the next level.
At the final level, matching results of 3-pixel grid points were
evaluated based on five test regions with different terrain types.
Region 1 is a relatively flat area at the Spirit rover landing
center. Region 2 is crater northeast side of Bonneville crater.
Region 3 is the summit of Husband Hill. Region 4 is the Inner
Basin area, located on the south side of the summit. Region 5 is
Home Plate.
Figure 4. Distribution of check points at five test regions
989