77ze International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
959 
3. NEW DEVELOPMENTS FOR AUTONOMOUS 
ROVER LOCALIZATION 
In our new autonomous rover localization approach, VO and 
BA methods are integrated with the expectation of achieving 
high efficiency and full automation. As illustrated in Figure 5, 
BA is performed at waypoints (panoramic sites and mid-point 
survey positions), while VO is performed between waypoints. 
The BA obtains the following data from VO: tracked features, 
refined image-orientation parameters as an input, and first and 
last stereo pairs. After BA, rover positions are updated at 
subsequent waypoints. The overall flowchart of this 
autonomous BA-based rover localization process is shown in 
Figure 6. This process includes initialization of image 
parameters (including inputs from VO), extraction and 
matching of interest points, selection of tie points, and bundle 
adjustment. 
Figure 5. Configuration of the onboard image network 
Figure 6. Flowchart of autonomous BA-based rover localization 
The key to the success of autonomous BA is selection of tie 
points, in particular, cross-site tie points. A great challenge is 
that the cross-site tie points can look significantly different 
from different viewpoints, especially from forward- and 
backward-looking images. We have developed a new approach 
to automatic selection of rocks as cross-site tie points through 
rock extraction, rock modeling and rock matching (Figure 7). 
Pre-screening and fault detection algorithms were also 
developed to ensure there is no mismatch in the final tie-point 
selection results. 
Input and output data 
Figure 7. Diagram of automatic cross-site tie-point selection 
Rocks are the major landmarks that can be easily identified in 
most of the ground images. Usually, rocks are composed of 
distinct rock peaks and surface points. A rock peak is extracted 
as the local maxima within a predetermined window from set of 
3D ground points generated by dense matching of stereo images. 
Starting from this extracted rock peak, a plane is estimated 
using those terrain points within an area of 70x70 cm from the 
rock peak. The initial rock height H is calculated as the 
perpendicular distance from the peak to the fitted plane. Surface 
points are searched for iteratively among the candidate points 
above the fitted plane using a dynamic search range of kH, 
where k varies from 0.3 to 1.7 based on a ground truth 
experiment in which manual measurements of rocks at the 
Spirit site were made and the coefficient k was calculated. 
Figure 8 shows examples of rock peaks and rock surface points 
extracted from Spirit rover images. The green dots are the rock 
peaks, while the red dots are the extracted surface points. 
Figure 8. Examples of extracted rocks showing peaks (green 
dots) and surface points (red dots) 
Each rock is then modeled using one of a number of analytical 
surface models such as hemispheroid, semi-ellipsoid, cone and 
tetrahedron. The parameters of each individual rock model for a 
rock are estimated by a least-squares fitting using the surface 
points on the rock. The model with the minimum root-mean- 
square error is considered the best model for that rock. 
Rock matching was used to find corresponding rocks in the two 
sets of rocks extracted from two different sites. The rock 
matching technique we have developed uses rock pattern 
matching to describe global rock-distribution patterns and rock 
model matching to depict individual rock similarities (Li et al.,