The SGM method itself uses a smoothness constraint that
penalizes neighboring pixels with different depth values. The
matching cost and the smoothness constraint are expressed in
a global cost function that is minimized by pathwise
approximation. SGM is not sensitive to the choice of
parameters, which is very important in practice since hand
tuning of parameter is avoided.
The inner processing loop of SGM is very regular and only
compares and adds integer values: This allows very efficient
CPU implementations using vector commands as well as
implementations on graphic cards (Ernst and Hirschmüller,
2008) and FPGAs (Gehrig et al., 2009; Hirschmüller, 2011).
SGM produces dense matching results in the resolution of the
input images (Figure 3). Typically, fine structures as well as
sharp object boundaries are precisely reconstructed.
Therefore, the method is very well suited for matching
airborne as well as satellite images (Gehrke et al., 2010;
Hirschmiiller and Bucher, 2010).
4. TESTS AND RESULTS
The presented method for automatic processing has been
tested with many different sets of stereo imagery from
QuickBird II (QB2), WorldView I (WV 1) and WorldView II
(WV02), kindly provided by DigitalGlobe. From many
successfully processed scenarios from 16 different sites of the
world, the results of four exemplary cases were selected to be
presented in detail.
The first scenario consists of two WorldView I stereo triples
of Berlin (Figure 4), an urban area with some small lakes and
soft elevations. Every triple, consisting of a forward, nadir
and backward scan, was taken within a few minutes from the
same orbit. Experience has shown that with this configuration
the best relative orientation between the images can be
achieved. Indeed, the relative accuracy over all images was
very high (Table 2), and the epipolarity error e, below half
the GSD. This means that the DSM can be generated at the
original resolution (GSD) of 0.55 m.
The second scenario is CapeTown (Figure 5), a mainly flat
urban area with high elevations in between and a coastline to
the sea. All nine images were captured at different seasons.
Therefore, the relative accuracy is much lower than in the
previous scenario. But at least a DSM with a GSDpsm of
2: e, = 0.83 m was achieved.
For the mountainous Dunedin scenario four stereo pairs were
used. Two of them were taken by WorldView I and two by
WorldView II. A similar accuracy as in as the Cape Town
scenario was achieved.
Scene Berlin 1 Cape Dunedin | Berlin 2
Town
1 WV1 2*2 WVI
* > , *
Sensors |2*3 WVI 8 WV2 2:2 WV2 2320 QB2
Area 500 km? 430 km? 580 km? 700 km?
GS Dax 10:55 m 0.55 m 0.57 m 0.63 m
€, 0.22 m 0.38 m 0.35 m 2.5m
€, 4.46 m 1.84 m 10.1 m 4.24 m
€, 0.24 m 0.43 m 0.39 m 1.86 m
tpreparation | 1.23 hours | 2.15 hours |0.88 hours | 0.25 hours
GSDpsm |0.55 m 0.76 m 0.70 m 3.72m
tsGM 40 hours 41 hours 87 hours 0.17 hours
Table 2: Results of the tests. e: The RMS of the spatial
distances of corresponding lines of sight of independently
selected homologous points (relative accuracy). e; The RMS
of the distances of the lines of sights from the GCP's absolute
positions (absolute accuracy). e,: epipolarity error.
Finally, another scenario from Berlin was chosen, consisting
of two Quickbird stereo pairs. All images were taken at
different times and seasons. Despite of that, the automatic tie
point selection worked very well. The achieved relative
accuracy was much worse than with imagery of the
WorldView satellites, also in other scenarios not included in
this paper. Probably this is due to the generally lower
geometrical precision of the QuckBird in compare to its
successor WorldView.
The achieved absolute accuracy (e,), as shown in Table 2, is
within the range of the absolute pointing accuracy of the
corresponding satellites. Both, the absolute and relative
accuracy have been checked with a set of precisely measured
and manually selected check points not used for bundle
adjustment in order to provide an independent reference.
The tests were performed on a Dell PowerEdge T610 with
two Intel Xeon X5570 Qudcore CPUs at 2.93GHz. For all
scenarios the manual preparation could be performed in at
most a half an hour. The automatic preparation time {preperation
needed until the SGM processing step requires relatively little
time when compared to the actual SGM processing step on
the same computer (see Table 2). The Dunedin scenario took
the longest to process, due to the larger height range, caused
by the mountains. Berlin 2 only took very little time as
GSDpsm Was chosen to be 3.72 m because of the low
accuracy of the relative orientation.
The time consumption for SGM processing per square
kilometer depends mainly on the GSDpsm, the number of
matches and the height range. In Table 2 the processing times
for the last two steps (SGM matching and orthophoto
generation) are given for the different scenarios.
Figure 4: Reconstruction of a part of Berlin, textured by
panchromatic images.
For :
and
half
expe
imag
It hz
surf
requ
pres
gain
num
with
deve
from
Ihe
can
para
matc
mori
The
Suit:
data
selec
refin
imag
even
Besi
As 0
1Ssuc
easil
are (
high
to us
Bay,
Spee
Und
Bou;
Kan:
Repc