The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part Bl. Beijing 2008
961
unable to select enough cross-site tie points (whether due to
lack of significant rocks or long traverse lengths greater than 30
m) the software was still very helpful in assisting operators to
rapidly select cross-site tie points for segments within 30 m.
The process was reduced to just several minutes. In the past, it
would take tens of minutes or even hours to manually select
only one cross tie point. This demonstrates that the software is
being effectively used in the ongoing MER mission for daily
operations.
4.2 Verification using Silver Lake Field Test Data
In order to test the performance of the autonomous BA
algorithm and the integration of BA and VO, a field test was
conducted at Silver Lake, California, on January 14, 15 and 16,
2007. A radio-controlled LAGR rover (Matthies et al., 2005),
capable of capturing panorama and VO images, traversed about
5.5 kilometers (as shown in Figure 12). VO images were taken
continuously at a rate of 2 fps; BA panoramic images were
taken at the ends of traverse segments with a typical segment
length of 25 m. The positions of the rover were obtained from
the DGPS (Differential GPS) at a data acquisition rate of 2 Hz,
which matched the VO image acquisition rate. The DGPS-
determined rover positions were used as ground truth to
evaluate the localization accuracy of BA and the integration of
BA and VO.
Figure 12. Silver Lake traverse (base map from Google Map)
Along the entire 5.5 km traverse, the rover acquired about
20,000 frames of VO images with a step length of 30 cm and
186 sites of panoramic images (3534 stereo pairs). Both the VO
and panoramic images are 1024x1024 pixels in size and 8 bits
in grayscale. The terrain captured in these images falls into
three categories: rocky outcrops, bushes, and dry lakebed (see
Figure 12). The stretch of rocky outcrops ran for about 210 m
and was imaged by panoramas from 14 sites. Panorama images
from 80 sites covered the bushy area along a traverse about 2.2
km. Although the shape of bushes is different from that of rocks,
which are the main features on Martian surface, our software
achieved a correct percentage of about 76. The remaining
images were mainly obtained on the dry lakebed. Cracks
running across the dry lakebed were very convenient for image
matching in VO, but make it impossible to pick cross-site tie
points for BA.
In order to evaluate the performance of both the BA and the
integration of BA and VO, we tested our software in two
different ways: without VO data, and with the integration of
VO data. Before the VO processing result was provided, the
software was tested only with the panorama data. The positions
of the panorama images were initially obtained by DGPS. They
were added with an artificially set noise of 10 percent of the
distance between two consecutive panorama sites. This 10
percent figure was used to make this positioning error
equivalent to the maximum positioning error from wheel
odometry. After the cross-site tie points were selected, the
image position and attitude (both having errors) are refined by
bundle adjustment. This test was conducted in an area of rocky
outcrops with a traverse of 14 segments (206 m). For these 14
segments, the software was able to automatically achieve
correct cross-site tie points within 11 segments. One segment of
31 m was excluded by pre-screening and 2 segments were
excluded by fault detection. The success rate, therefore, is 79
percent (85 percent after pre-screening).
Computation of the VO and BA integration was performed in
the local coordinate frames (east-north-up) of three stretches of
traverses. Each local frame origin is at the center of the first
panorama of that stretch. The orientation (azimuth and tilt) of
the first panorama in each local frame was calculated manually
based on the first panorama, an adjacent panorama for this
stretch of traverse, and DGPS data. We matched the first VO
pair to the first panorama (already in the local frame) and then
transformed all the VO poses to the local frame. After BA, we
evaluated the localization error at the end point by comparing
the BA-derived position and DGPS position (as shifted to the
local frame). Figure 13 shows the BA and VO integration
results in the rocky outcrop area. The blue, red and blacks lines
represent rover traverses from VO, from integration of VO and
BA, and from ground truth, respectively. We can observe that
the integrated BA and VO were significantly better than the
initial VO result. This indicates that the BA panoramas
improved the geometric strength of the image network and
provided better localization accuracy than VO alone. Relative
accuracy improved from 27.1 to 3.9 percent. As for automatic
cross-site tie-point selection, we obtained the same success rate
as results without VO data in the rocky outcrop area.
Figure 13. Result of integration for rocky outcrop area near
Silver Lake, CA (units: m)
We also tested the software in a bushy area around Silver Lake
using a total of 81 pairs and a traverse length of 2.2 km. Among
the 81 sites were 10 consecutive pairs (about 0.7 km) whose
traverse lengths were approximately 50 m each, which is
beyond our software’s ability to reliably extract features. The
fault detection module excluded these 10 pairs, thus they are