«1 2004 
  
ft); 
ight) 
(1) 
S 
I 
) image 
de) 
de) are 
"ure 6. 
  
els of 
nced by 
points. 
luce the 
Instead, 
> a over 
1e mean 
iting its 
norama 
ight). 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 
  
30030001 -adjust.bmp 380030041-adjust.bra 
  
350030071 .bmp 350030081 .bmp 
50030061 -adjust.bmp 350030071-adjust.bm 
Figure 7. Intensity balancing using tie points: original i images 
(left) and adjusted images (right) 
34 DEM Interpolation and Orthophoto Generation 
From the first panorama of each landing site, we extracted 
interest points and calculated the statistics of the Martian terrain: 
semivariogram. We found that a dual polynomial model, one 
for close range (0~5m), another for far range (5-50m) fits well 
with the semivariogram, as shown in Figure 8. 
    
Figure 8. Dual polynomial model for Kriging: close range 0-5m 
(left) and far range > 5m (right) 
The distribution of 3-D interest points is unbalanced: it is dense 
in the close range and sparse in the far range. Normally in the 
very far range (> 25m) of Navcam, the Kriging model will no 
longer work, so we used both Kriging (in the range < 25m) and 
TIN (in the range > 25m) to interpolate the DEM. 
  
Figure 9. DEM and orthophoto for MER-B Site 14 (Fram Crater) 
4. ROVER LOCALIZATION 
The photogrammetric solution to rover localization uses the 
output of the rover’s s navigation sensors as an initial value and 
improves its precision from ten to one percent (Li, 2002). 
The key to improvement of precision is to find sufficient tie 
Points between cross-site image pairs. Since different sites are 
normally separated by over 50 meters (the length of a sol's 
travel), usually only obvious landmarks like rocks can be 
identified. 
1045 
  
where 
  
, Vol XXXV, Part B4. Istanbul 2004 
4.1 Landmark Extraction 
Landmarks (most often rocks on Martain surface) can be 
detected from occlusion. In ground level rover images, a rock 
normally occludes a long region behind it, see Figure 11. The 
elevation of these occlusions, however, can be interpolated 
correctly via Delauney triangulation of interest points. Thus, by 
projecting the DEM to the image plane and comparing the 
calculated parallax with actual parallax, occlusion can be 
detected, which reveals the size and shape of the front side of 
the rock. 
  
Figure 10. 3-D meshed-grid of the Fram Crater 
  
Figure 11. Interest points; DEM; parallax difference; and 
detected rock occlusion 
Rocks can then be modelled as a half ellipsoid with an 
uncertain backside, as seen in Figure 
12 (right), the measurement uncertainty of the rock can be 
modelled as an ellipse with parameters: 
12 (left). As seen in Figure 
2 
a= s. dp 2 
bf I] (2) 
D = sA 
a, b are the long / short half axis 
dp is the parallax error, around 1/3 pixel 
b is the baseline length 
f is the camera focal length 
A is the angular resolution of camera 
Figure 12. Rock model (left) and its measurement uncertainty