1167
V* »
' •
- \ s • -
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
matching is repeated. The whole procedure must be repeated
until the last layer of the image pyramid is reached. For the
estimation of subpixel positions, a least-squares matching is
performed in the last layer given the preliminary matching
points.
In the next step, a region-growing matching (Heipke, C., 1996;
Otto, G.P. 1989) based on the results of the hierarchical
matching is performed. The region-growing must be performed
in three steps, because dense tie points in three images must be
found. First, the standard bidirectional stereo region-growing is
performed on the nadir-backward image pair using the tie points
from the hierarchical matching as seed points. Then, the
resulting tie points are searched in the forward image. Finally,
the resulting 3ray tie points are cross-checked with a matching
from forward to the backward image.
e.g. 1:18 and a flight height of 2000m acceptable accuracies of
1.33m could be reached, which are sufficient for some
applications in the context of disaster monitoring.
3. PERFORMANCE RESULTS
The DSM generation processing flow was validated with 3K
image data from Munich at two different flight dates. On 30 th
April 2007 two flight strips from 1000m resp. 2000m a. G., and
on 17 th June 2007 flight strips from 2000m were acquired from
the centre of Munich.
SM
Given the tie points, the image positions and attitudes, as well
as the calibration data, the 3D coordinates of tie points are
calculated by forward intersection. Outliers could be detected in
different ways. First, they are eliminated by the ratio
residue/G residues of the image coordinates in the least-squares
adjustment. An advantage of using 3ray points is that outliers in
the matching could be better detected than by using stereo
points due to the higher redundancy. Second, outliers are
filtered out by using the standard deviation of the resulting
coordinates as well as by applying a value range rule.
2.2 Simulation study
Table 1 shows the expected height accuracies a z based on a
simulation of different flight and image acquisition
configurations. Accuracies for two flight heights above ground
(1000m and 2000m) and for four different base-to-height ratios
B/Z from 1:5 to 1:18, were simulated based on a varying
number of images (3ray points resp. 5ray points). Given a
specific flight and image configuration, the required frame rate
in Hz can be derived. For the simulation, the relative accuracies
of image positions and attitudes were assumed to be very high,
whereas systematic deviations were not taken into account. The
matching accuracy was set to 0.15 pixels.
Him]
B/Z
ct z (3 images)
Hz
a z (5 images)
Hz
2000
1:5
0.34m
0.5
0.30m
0.8
2000
1:9
0.66m
0.9
0.59m
1.5
2000
1:18
1.33m
1.6
1.14m
2.7
2000
1:36
3.33m
3.6
2.97m
5.9
1000
1:9
0.36m
1.6
0.31m
2.7
1000
1:18
0.88m
3.6
0.80m
5.9
Table 1 Expected height accuracy for different configurations
The results show - as expected - best accuracies for higher
base-to-height ratios and for lower flight heights, e.g. when
using 3ray points from 2000m a. G. and a base-to-height ratio of
1:5 the standard deviation is 0.34m. In general, the standard
deviation improves only by ten percent if 5ray points are used.
Nevertheless, higher base-to-height ratios cause stronger
“image decorrelation” in particular in urban areas, i.e. the
number of matched points increases for smaller base-to-height
ratios, as e.g. similar building sides are imaged. Smaller base-
to-height ratios increase on the other hand the frame rate, which
could be only achieved by new digital cameras like the Canon
EOS Mark II. Besides, even with a small base-to-height ratio of
tom?
Figure 5
3K-DSM from
Nymphenburg
the center of Munich: Castle
The base-to-height ratio for the generation of DSM was quite
small (1:17 resp. 1:9) because images were acquired as bursts at
1.6 Hz. On 17 th June 2007, the same area was imaged from
2000m a. G. Here, the base-to-height ratio was higher (1:4), as
images were acquired continuously with a low frame rate of
0.5 Hz.
3K DSMs from a reference region were generated for all flight
strips according to the proposed processing flow (see example
in Figure 5). All 3K DSMs were compared with a high
resolution reference DEM based on LIDAR measurements.
H[m]
B/Z (Hz)
a z (sim.)
RMSE (empir.)
30/04/2007
2000
1:17(1.6)
1.33m
1.91m
30/04/2007
1000
1:9 (1.6)
0.36m
0.35 m
17/06/2007
2000
1:4 (0.5)
0.34m
0.85m
Table 2 Accuracies of 3K DSM over flat terrain
Table 2 lists the simulated and empirically tested height
accuracies of the processed 3K DSMs over flat terrain without
vegetation and buildings. For the empirical determination of
accuracies, height differences in small control areas were
measured, which are not vegetated and are mostly flat.
The simulated accuracies from 1000m a. G. (30/04/2007)
correspond quite well with the empirically derived accuracies,
whereas in the other flight strips the empirically derived ones
were higher than in the simulation. One reason for the deviation
are systematic errors, e.g. in the measured GPS coordinates,
which are not included in the simulation, but produce
systematic offsets in the generated 3K DSMs.
Small systematic effects could also be detected in figure 6,
where histograms of the difference images, the 3K DSM minus
the reference DEM, were plotted for the three data takes.
Ideally, the histogram has a sharp peak at the difference zero,
which could be observed e.g. in the histogram for the data take
on 30 th April from 1000m. But in the other difference