The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
Non Local Means (Buades et al. 2005) are not able to yield
advantage over a simple Gaussian. Only with stronger filtering
the edge preserving nature of these algorithms becomes
effective. However, for tie point extraction minimal filtering
with a 3x3 pixel Gaussian window already generates on
average about 50% more points compared to unfiltered images.
MOLA DTM
Exterior orientation
>
HRSC
level-2 images
▼
opt. low pass filtering
▼
Filtered
level-2 images
3L_
Prerectification
v
Rectified
level-3 images
m *
Image matching Level-2 coordinates
▼
Image coordinates
level-3
▼
Calculation of level-2 coordinates-^
i
Image coordinates
level-2
Figure 1. Processing chain of tie point extraction
In the next step the images are prerectified (Norvelle 1992,
Scholten et al. 2005) using the exterior orientation provided by
ESOC and the MOLA DTM. The resulting ortho photos (called
“level-3 images”) serve as input for the matching. The
prerectification is necessary to compensate for scale differences
caused by the elliptical orbit and for non-quadratic pixels
caused by not perfectly adjusted integration times of the line
scanner. As an example of the geometric differences which can
occur a section from orbit 30 is shown in Figure 2:
Stereo 1 channel Nadir channel Stereo 2 channel
Level-3 image of the nadir channel
Figure 2. Geometric differences in images of the same strip
In the first row the crater is shown in the original level-2 images
of the forward, nadir and backward looking channel.
Underneath, the same area is shown in the rectified version of
the nadir channel. During rectification information about the
origin of the level-3 pixels with respect to the level-2 images is
stored in separate files (labelled “Level-2 coordinates” in Figure
1). After the matching this information is used to convert the
level-3 image coordinates of the tie points back into the level-2
coordinate system of the original images where the collinearity
equations used by the bundle adjustment are defined.
The matching employs a pyramidal approach to account for big
parallaxes and imprecise approximate values of the exterior
orientation. Generally, the nadir channel is matched pairwise
with the remaining four panchromatic channels called stereo 1,
stereo 2, photometry 1 and photometry 2. Additionally, it is
possible to use the four colour channels blue, green, red and
infrared if they are available at a decent resolution. Due to a
low bandwidth between Earth and Mars for data transmission
their resolution is often reduced via pixel binning (macro pixel).
For example, a macro pixel format of 8 x 8 pixels decreases the
size to l/64th compared to the original amount of data which is
inadequate for precise point determination.
The candidate points are distributed in a grid structure over the
master image (nadir channel). Practical experience has shown
that this approach is advantageous over a feature based operator
with respect to HRSC data. This allows a more complete
coverage with tie points and a more accurate adjustment of the
point cloud to the MOLA DTM. An interest operator
preferably generates points at edges which often coincide with
breaklines in HRSC images. Because of the lower resolution of
the MOLA dataset the differences to the HRSC points at these
locations are higher than in flat terrain. Point transfer from the
master image to the candidate image is carried out using the
cartographic coordinates of the level-3 images.
The Normalised Cross-correlation Coefficient (NCC) is used as
similarity measure which can compensate for radiometric
differences with respect to gain and contrast. The NCC
assumes parallel image and object planes without elevation
differences in the correlation window. To a large extent this
assumption is fulfilled in the prerectified images. Therefore, it
is possible to use large correlation windows of 35 x 35 pixels or
bigger which are needed to capture a sufficient number of grey
value variations in low texture imagery.
To increase the accuracy of the points a multi-image least
squares matching (MI-LSM) is carried out in which all points of
a tuple are matched simultaneously. Following the approach of
Krupnik & Schenk (1997) it is not necessary to estimate six
affine transformation parameters per patch. As a result of the
prerectification it is sufficient to estimate two horizontal
translations only. The MI-LSM minimises the grey value
differences between a particular image patch g', /=0,...,n
(n+1 = number of image patches) and the theoretical grey
values of the surface patch (reference patch). Ideally the
differences are zero but in reality errors occur because of
radiometric and geometric inaccuracies:
G(x,y)- g' (x + dd',y + db‘)= v(x,j>) (1)
where G(x,y) = theoretical grey values of reference patch
g'(x, y) = image patch grey values
da' ,db‘ = translations of patch i
v = residuals
1026