The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
960 
2007b). The peaks of the matched rocks that pass both rock 
pattern matching and rock model matching are taken as cross 
site tie points. To ensure the effectiveness of the matching, and 
an even distribution of the tie points, we define a 4x4 grid in the 
area of overlap between the two sites. Within each grid cell, we 
select a limited number of significant rocks (e.g., up to 3), 
which are usually the highest rocks in the grid cell. Only the 
selected significant rocks at the current site are used to find 
their corresponding rocks at the adjacent site. Figure 9 shows an 
example of automatically selected cross-site tie points between 
Sites 1200 and 1300 of Spirit rover. These two sites are 23 m 
apart and the image data were acquired on Sols 61 and 62, 
respectively. 
Figure 9. Automatically selected cross-site tie points between 
Sites 1200 and 1300 of Spirit rover 
Pre-screening and fault detection were based on extensive tests 
and statistical analysis. We found that traverse distance, 
distance ratio, and the number of used peaks were the most 
important factors for fault detection. The distance ratio 
compares two distances: a rock to the camera position at on site 
versus the same rock to the camera at the adjacent site. At the 
pre-screening step, pairs with the following conditions were 
excluded: 1) traverse leg length being less than 30 m, or 2) 
number of rock peaks extracted being less than 20, which is not 
sufficient for significant peak selection. In fault detection, we 
excluded rocks with distance ratios less than 0.3, rocks with 
unreliable modeling parameters, rocks with unmatched local 
terrain at both sites, and sites whose number of matched rocks 
in the final result were less than 3. These pre-screening and 
fault detection strategies ensure that the successfully selected 
cross-site tie points are of high quality. 
4. VERIFICATION OF AUTONOMOUS ROVER 
LOCALIZATION TECHNOLOGE USING SPIRIT 
ROVER DATA AND FIELD TEST DATA 
4.1 Verification using Spirit Rover Data 
We have tested our new software using a 318 m traverse (19 
pairs of sites) taken by Spirit from Sols 574 to 648 in the 
Husband Hill summit area. The test results are shown in Figure 
10. Black dots show sites where Navcam or Pancam panoramic 
images were taken. Green segments delineate traverse legs 
outside of the test area. Red segments designate those traverse 
segments that have passed fault detection and successfully 
completed BA. Both yellow and orange segments show those 
segments that failed the BA. Yellow segments designate those 
excluded by pre-screening, while orange segments designate 
those excluded by fault detection. A success rate of 68 percent 
(13 out of 19 pairs) was achieved, which is a very successful 
result considering that the MER-A traverse was not designed 
originally for autonomous BA. 
The performance of the autonomous BA is shown in Figure 11, 
where differences between the blue (telemetry-based) and red 
(BA-based) lines represent the differences between the 
telemetry and BA positions. Traverse segments that passed fault 
detection and successfully finished BA are indicated by solid 
red lines while the dashed red lines represent those pairs of sites 
that failed BA, whether excluded by pre-screening or fault 
detection. Test results using MER data have shown that the 
proposed method is effective for medium-range traverse 
segments (up to 26 m). As an example, in the first segment 
(Sites 11304 to 11308), BA corrected the rover’s position by 
5.6 percent (0.95 m out of a total segment length of 16.96 m). 
Figure 10. Map of the bundle-adjusted rover traverse of the 
Spirit rover in the Husband Hill summit area 
Figure 11. Comparison of Spirit rover traverses in the Husband 
Hill summit area. Blue line is the telemetry-based traverse and 
the red line is the traverse computed from the autonomous BA. 
Since August 2007, this newly developed software has been 
employed to perform automatic rover localization for the Spirit 
rover in the Home Plate area (Figure 1) in the Earth-based data 
processing environment for ongoing MER mission operations. 
The developed software has been able to automatically select 
cross-site tie points for 71 percent of the total number of 38 
traverse segments. Over a traverse of 270.92 m, it has corrected 
the rover’s position by 11.03 m (4.07 percent). For the 
remaining 29 percent of the traverse segments, despite being