The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
3 EXPERIMENTS 
The proposed workflow is demonstrated using two example video 
datasets. The first dataset UAV was obtained from a Microdrone 
(Microdrones, 2008) platform and captured near the ’’Drachen- 
fels” close to the city of Bonn (Forstner and Steffen, 2007). The 
second dataset (FLI-MAP video) was captured from a helicopter 
during a LIDAR flight over Enschede (Fugro, 2008), see Table 1 
for some parameters. In that table, the baselength refers to two 
Parameter 
UAV 
FLI-MAP 
Flying height H (m) 
30 
275 
Image scale 1 : mt, 
1:1,500 
1:50,000 
Frame size (pix) 
848x480 
752x582 
Pixel size (//m) 
12 
8.6 
Frame rate (Hz) 
30 
25 
Approx, baselength b (m) 
0.1 
1 
Length sequence (img) 
280 
150 
Table 1: Some parameters from the datasets. 
consecutive frames. The length of the sequence refers to the num 
ber of images which were used for the examples. Noteworthy is 
the small image scale from the FLI-MAP video; the calibrated fo 
cal length of this video device is only 6mm. From this geometric 
set-up no highly accurate forward intersection can be expected, 
refer to the section on the resulting point cloud. Some undistorted 
images from both sequences are shown in Figure 4. 
3.1 Results 
3.1.1 Super resolution images An example for a super reso 
lution image is taken from the UAV dataset. The chosen target 
scale factor is 1.5. In Figure 5 the gray value profiles (red chan 
nel) across the building’s roof edge are shown. The left image 
shows in its upper area a part of the original image, but scaled by 
factor 1.5 (linear interpolation applied). The line across the edges 
indicates the location of the grey value profile as shown below the 
image. The SRI image, computed from the mean value of corre 
sponding points is shown in the right part of Figure 5, including 
the gray value profile captured at the same image position as in 
the original image. 
In general one can see that the SRI seems to be a bit sharper com 
pared to the original one: The tiles on the roof are less smeared 
than in the original image. The profile supports the visual impres 
sion. Especially in the edge region more details are shown. As an 
example two points at the profile graph are pointed out by a black 
arrow. The arrow no. 1 points to the quite salient point in the 
profile indicating the position of the steep edge where the light 
grey becomes dark grey. The corresponding area in the profile 
of the original image is smoother. The arrow no. 2 points to the 
edge at the eave of the roof where the tiles are showing a lighter 
colour compared to the red colour on the overall roof area 1 . In 
the computed SRI image this edge is really existing, but not in 
the original image. 
The SRI as computed from the respective median value of cor 
responding pixels is not shown here, because no significant dif 
ference can be observed compared to the SRI computed from the 
mean value. This can be explained by the use of solely redundant 
matches: by this means no gross errors are expected which may 
influence the SRI from the mean values and thus the robust values 
from the median computation are close to the mean. 
3.2 3D point cloud 
In order to evaluate the expected accuracy from forward inter 
section first a theoretic approximation for the height accuracy is 
made. In the given examples, especially in the FLI-MAP video, 
the height component is the critical component. 
Generally, given an approximated stereo normal case, the height 
difference between two points is estimated as 
If only the uncertainty in parallax measures s px is considered, 
the accuracy for height measurements is derived from the partial 
derivative with respect to p x : 
In the case at hand more than 2 rays are intersected and thus the 
expected accuracy and reliability is higher, but nevertheless the 
approximation reasonably reflects the quality for forward inter 
section.