In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009 
Parameter 
FLI- 
vertical 
VIAP 
oblique 
Pictometry 
oblique 
flying height [m] 
275 
275 
920 
baseline [m] 
50 
50 
400 
tilt angle [°] 
0 
45 
50 
number of images,viewing di 
rection 
7 
8xSW 
2xW, 2xS, 
2xE, lxN 
focal length [mm] 
35 
105 
85 
pixel size [p,m] 
9 
9 
9 
sensor size [mm x mm] 
36x24 
36x24 
36x24 
GSD and theoretic accuracies (for oblique: from fore- to background) 
ground sampling distance [cm] 
7 
2.8-4 
10- 16 
sx,y, vertical [cm] 
4 
NA 
NA 
sz, vertical [cm] 
40 
NA 
NA 
sx,y, across track base [cm] 
NA 
NA 
22-44 
sz, across track base [cm] 
NA 
NA 
18-37 
sx,y, along track base [cm] 
NA 
60-92 
22 - 42(*) 
sz, along track base [cm] 
NA 
60-92 
19 - 35(*) 
(*): along track base images from Pictometry were not used. 
Figure 2: Image parameters and layout of sample block 
3.2 Rectification and dense matching 
The approach to dense stereo matching as applied in the current 
implementation is the Semi-Global-Matching algorithm (Hirsch- 
miiller, 2008). The basic idea behind this technique is to aggre 
gate local matching costs by a global energy function, which is 
approximated by an efficient pathwise 1-dimensional optimiza 
tion. 
To simplify the matching, the images are rectified beforehand. 
For this purpose the approach proposed in (Oram, 2001) is ap 
plied. A homography is estimated which is compatible to the fun 
damental matrix (Hartley and Zisserman, 2004, chap. 13). The 
aim is to minimize distortions due to perspective effects, and thus 
also to reduce the disparity search space. A consequence from the 
particular algorithm is that the epipolar lines are coincident, but 
not necessarily parallel. Hence, in the subsequent rectification the 
images need resampling to obtain parallel epipolar lines. One dis 
advantage of this procedure is that straight lines are not preserved, 
however, this does not influence the matching. To compute the 
optimal homography point correspondences are required, like for 
instance in the case at hand the adjusted tie points. If images are 
taken from approximately the same viewing direction, it is also 
possible to extract further matches through scale invariant point 
descriptors like SIFT (Lowe, 2004). Outliers in the correspon 
dences are identified through RANSAC within the estimation of 
the compatible homography. The inliers are here also used to 
estimate the disparity search range for the dense matching. 
4 EXPERIMENTS 
4.1 Description of used data 
Part of the data used for these experiments was acquired by the 
Fugro Inpark FLI-MAP 400 system in March 2007 over Enschede, 
The Netherlands. Besides two LIDAR devices and two video ca 
meras, the system carries two small-frame cameras, one pointing 
vertical, and one oblique camera, looking in flight direction, tilted 
by approx. 45°. Additional Pictometry images were made availa 
ble through BLOM Aerofilms. Those images were acquired only 
one month before the FLI-MAP data. A small block of 7 ver 
tical and 8 oblique images from FLI-MAP as well as 7 images 
from Pictometry was chosen for the experiments. In Fig. 2, upper 
part some parameters of the images are given, the GSD and ac- 
curay estimation was done according to equations 1 to 11, while 
a standard deviation for image measurements of a half pixel was 
assumed. In the bottom of that figure the layout of the block is 
shown, including GCP, check points and the approximate posi 
tion of defined scene constraints. The highly overlapping ima 
ges in the center are from the FLI-MAP acquisition, while the 
7 regularly aligned outer images are from the Pictometry-flight. 
Note that no along track images are chosen from Pictometry. The 
airplane acquired the images in N-S-direction, so the East- and 
West-looking images belong to one flight line (baseline approx. 
400m) and the two South-looking images are from two adjacent 
strips, baseline approx. 350m. For the accuracy estimation the 
two South-looking images can be treated like across-track ima 
ges. 
4.2 Block adjustment results 
Four full and one height GCP were used for the adjustment. Ad 
ditionally, one right angle, 3 horizontal and 4 vertical line con 
straints were defined. It was assured that in every image at least 
one of the features used for the scene constraints was visible. In 
Table 1 the adjustment results in terms of RMSE at the control 
and check points, or features respectively are listed. One obser 
vation from the residuals is that the Z-component is smaller than 
the X/Y values for all control and check features. Also the resid 
uals at vertical constraints are larger than the residuals at horizon 
tal constraints, and those are also influenced by the Z-component 
only. One reason for this can be that the tilt of the Pictometry 
images is larger than 45° and thus the X,Y-component is less ac 
curate than the Z-component, refer also the the listed theoretic 
accuracies in Fig. 2. One general drawback of this block-setup is 
that outside the overlapping areas no GCPs or scene constraints 
are available and applicable, respectively, so the overall block ge 
ometry at the borders is not optimal. However, since the residuals 
at the façades are at least for the Pictometry images less than one 
pixel this result can be considered satisfactory. 
Assessment 
RMSE value[cm] 
X-Res. at GCP 
2.1 
Y-Res. at GCP 
4.8 
Z-Res. at GCP 
1.3 
X-Res. at Check 
16.2 
Y-Res. at Check 
5.8 
Z-Res. at Check 
1.5 
Res. at H-constraints 
1.4 
Res. at V-constraints 
6.8 
Res. at RA-constraints (°) 
0.01 
Table 1 : Residuals from bundle block adjustment 
4.3 Dense matching results 
The dense matching was performed in several stereo image com 
binations. Besides the matching in images from one platform,