The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
1138 
2.1 Stereo Matching 
Hierarchical intensity based matching is used for matching the 
stereo pairs and the reference image. It consists of two major 
steps, hierarchical matching to derive highly accurate tie points, 
followed by a region growing step to generate a dense set of tie 
points. 
The initial matching step uses a resolution pyramid 
(Lehner&Gill, 1992; Komus et al., 2000) to cope even with 
large stereo image distortions stemming from carrier movement 
and terrain. Large local parallaxes can be handled without 
knowledge of exterior orientation. The selection of pattern 
windows is based on the Foerstner interest operator which is 
applied to one of the stereo partners. For selection of search 
areas in the other stereo partner(s) local affine transformations 
are estimated based on already available tie points in the 
neighborhood (normally from a coarser level of the image 
pyramid). Tie points with an accuracy of one pixel are located 
via the maximum of the normalized correlation coefficients 
computed by sliding the pattern area all over the search area. 
These approximate tie point coordinates are refined to subpixel 
accuracy by local least squares matching (LSM). The number 
of points found and their final (subpixel) accuracy achieved 
depend mainly on image similarity and decrease with 
increasing stereo angles or time gaps between imaging. The 
procedure results in a rather sparse set of tie points well suited 
for introduction into bundle adjustment and as an excellent 
source of seed points for further densification via region 
growing. 
The second step uses the region growing concept first published 
by Otto and Chau in the implementation of TU Munich (Heipke 
et al., 1996). It combines LSM with a strategy for local 
propagation of initial conditions of LSM. Epipolar stereo 
images are used in the region growing step, and the propagation 
strategy is modified to enforce points located on the epipolar 
lines. Stereo tie points deviating more than 0.5 pixels from the 
epipolar geometry are removed. A quasi-epipolar stereo pair 
with epipoles corresponding to the image columns is generated 
by aligning the columns of the Fore image with the Aft image, 
using highly accurate matches from the pyramidal matching 
step. 
Various methods for blunder reduction are used for both steps 
of the matching: 
• Threshold for correlation coefficient 
• 2-directional matching and threshold on resulting 
shifts of the coordinates 
• Threshold on the deviation from epipolar geometry. 
In areas of low contrast the propagation of affine transformation 
parameters for LSM in region growing leads to high rates of 
blunders. In order to avoid intrusion into homogeneous image 
areas (e.g. roof planes without structure) the extracted image 
chips are subject to (low) thresholds on variance and roundness 
of the Foerstner interest operator. This and the many occlusions 
found in densely built-up areas imaged with a large stereo angle 
create lots of insurmountable barriers for region growing. Thus, 
for high resolution stereo imagery the massive number of seed 
points provided by the matching in step one (image pyramid) 
turns out to be essential for the success of the region growing. 
The numbers of tie points found and their subpixel accuracy is 
highly dependent on the stereo angle. A large stereo angle 
(large base to height ratio b/h) leads to poorer numbers of tie 
points and to lower accuracy in LSM via increasing 
dissimilarity of (correctly) extracted image chips. 
2.2 GCP collection and affine RPC correction 
Previous studies (Lehner et al., 2007) have shown that the 
CARTOSAT-1 RPC ground accuracy is in the order of hundred 
meters. Additionally, forward intersection performance without 
RPC correction is poor and results in large residuals in image 
space. The estimation of affine RPC correction parameters 
requires well distributed GCP with subpixel accuracy. In many 
application scenarios, such as continent wide reconstruction or 
crisis support applications, acquiring the required GCP is very 
tedious or might even be impossible, if a fast response is 
required. 
Gobal and easily available reference datasets are the OnEarth 
Landsat ETM+ Geocover mosaic and the SRTM elevation data. 
The accuracy of these datasets is low compared to the high 
resolution CARTOSAT-1 images. The Landsat ETM+ 
Geocover mosaic is specified with a lateral error of 50m. The 
absolute lateral error of SRTM amounts to 7.2m - 12.6m (LE90, 
depending on the continent), with an absolute height error of 
4.7m to 9.8m (Rodriguez et al., 2005). 
GCPs are collected by transfer of highly accurate tie points 
between the CARTOSAT-1 Aft and Fore images to the Landsat 
reference image and extraction of the corresponding height 
from SRTM. The matching procedure starts by aligning the 
CARTOSAT-1 Aft image to the ETM+ reference by using the 
comer values provided in the CARTOSAT-1 metadata. The 
first step of the hierarchical matching procedure described in 
Section 2.1 is applied to obtain tie points between the ETM+ 
and CARTOSAT-1 Aft scenes. A similar matching could be 
done by matching the Fore image against the ETM+ image, it 
would however yield different tie points and thus GCPs for the 
Aft and Fore image. Since affine RPC correction is performed 
separately for each image, a good link between the Aft and Fore 
images is required to ensure good forward intersection 
behaviour. Thus, highly accurate stereo tie points between the 
CARTOSAT-1 Aft and Fore images are selected by applying 
strict thresholds on the bidirectional matching shift (0.1 pixels) 
and correlation coefficient (0.8). The Aft coordinates of these 
stereo tie points are then used as interest points and matched 
against the full resolution ETM+ scene, using the previous Aft 
vs. ETM+ matching as initial approximation. This yields the 
geographic position of the stereo tie points. Finally, 3D GCPs 
for both Aft and Fore scene are obtained by bilinear 
interpolation of the SRTM DSM. We use a non-interpolated C 
band SRTM, where holes larger than 2 pixels are still open. 
This avoids deriving GCP from interpolated heights. Affine 
RPC correction parameters are estimated both for the Aft and 
Fore scene. 
2.2.1 RPC correction by DSM alignment 
After the alignment based on ETM+ and SRTM reference data, 
forward intersection residuals are significantly improved, but 
the lateral accuracy is still limited by the ETM+ Geocover 
reference. To take advantage of the higher accuracy of the 
SRTM dataset a second RPC correction step is necessary. A 3D 
point cloud is calculated by forward intersection of a subset of 
the stereo tie points. The point cloud is aligned to the SRTM 
DSM. It is assumed that the height z, of a point P, located at 
(Xj.yi.z,) equals the reference DSM height h n (x t ,y{) at the 
corresponding position (x„y,):