In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
DENSE MATCHING IN HIGH RESOLUTION OBLIQUE AIRBORNE IMAGES
M. Gerke
International Institute for Geo-Information Science and Earth Observation - ITC, Department of Earth
Observation Science, Hengelosestraat 99, P.O. Box 6, 7500AA Enschede, The Netherlands, gerke@itc.nl
KEY WORDS: Adjustment, Bundle, Calibration, Matching, Point Cloud, Rectification
ABSTRACT:
An increasing number of airborne image acquisition systems being equipped with multiple small- or medium size frame cameras are
operational. The cameras normally cover different viewing directions. In contrast to vertical images, those oblique images have some
specific properties, like a significantly varying image scale, and more occlusion through high raising objects, like buildings. However,
the faces of buildings and other vertically extended objects are well visible and this is why oblique images are used for instance for
visualization purposes.
This paper shows results from applying the sophisticated Semi-Global-Matching technique to a set of oblique airborne images. The
images were acquired by two systems, namely FLI-MAP 400 (Fugro Aerial Mapping B.V.) and Pictometry (BLOM Aerofilms) over
the same area. After the joint adjustment of the images, dense matching and forward ray intersection was performed in several image
combinations. The disparity maps were evaluated through the comparison with a reference map derived from LIDAR which was
acquired in parallel with the FLI-MAP system. Moreover, the 3D point clouds were analyzed visually and also compared to the
reference point cloud. Around 60 to 70 percent of all matches were within a range of ± 3pix to the reference. Since the images were
acquired in different flight configurations, the impact of different intersection angles and baselines to the triangulation is quite obvious.
In general, the overall structures on the building faces are well represented, but the noise reduction needs further attention.
1 INTRODUCTION
An increasing number of airborne image acquisition systems are
operational (Petrie and Walker, 2007). Because of the availability
of low-cost digital cameras with small or medium sized sensors,
some of those systems carry multiple cameras covering differ
ent viewing directions. For instance from Pictometry 1 image are
available already for a number of cities and they are accessible in
the category ’’birds eye view” in Microsoft Bing Maps 1 2 (formerly
known as Virtual Earth).
The use of oblique images for topographic mapping purposes was
shown in quite some papers. In (Hôhle, 2008) height determina
tion from single oblique images is demonstrated. The verifica
tion of vector data using oblique imagery is shown in (Mishra
et al., 2008). Due to the fact that building façades are well vis
ible in oblique images, some researchers concentrate on how to
automatically extract façade textures (Früh et al., 2004, Wang et
al., 2008). Besides, the oblique images are interesting for cadas
tre applications, because the building outline as defined at the
vertical wall is directly visible (Lemmen et al., 2007). Com
pared to vertical airborne images, oblique images have some spe
cific properties. Depending on the tilt angle, the scale within
the imaged scene varies considerably. Moreover, vertical struc
tures of raised objects like buildings or trees are imaged, but the
(selfjocclusion by those objects is much more significant com
pared to the vertical image case.
Another interesting application and research domain concerns the
derivation of high dense point information through image mat
ching techniques. The benchmark results from the Middlebury 3
testsets show that high quality state-of-the-art techniques to dense
matching are available. If it is possible to apply those techniques
to oblique airborne images, interesting new applications would
arise, or support existing ones, like the ones listed above. In gene
1 http://www.pictometry.com
2 http://www.bing.com/maps
3 http://vision.middlebury.edu/stereo/ (accessed 15 March 2009)
ral, point clouds as derived from dense matching in oblique ima
ges can be a complementary data source to airborne laser scan
ning, as those devices normally do not capture dense points on
vertical structures. Of course, the traditional use of this kind of
data to produce digital surface or terrain models is another possi
ble application.
In (Besnerais et al., 2008) an approach to dense matching in obli
que airborne images is presented. The authors develop a pixel
wise similarity criterion which accounts for the special viewing
geometry of oblique images. A dense depth map is obtained
through global regularization. The approach was tested on a num
ber of test images and showed good results. However, the ground
sampling distance of the used images was not smaller than 1,4m,
mostly it was even larger, up to 20m.
This paper evaluates the application of the Semi-Global-Matching
technique (SGM, see (Hirschmtiller, 2008)) to a set of high reso
lution FLI-MAP 4 and Pictometry images. One particular façade
of a building is normally only visible in images taken from one
viewing direction, resulting in a relatively bad intersection angle
in object space. Thus, the main objective of this paper is to eval
uate the overall accuracy of the derived 3D point cloud as derived
from a forward intersection of matched points. Although other -
may be better performing- algorithms for dense matching exist
(Seitz et al., 2006) we chose SGM, because it demonstrated al
ready its fitness for the photogrammetric production process, c.f.
(Hirschmüller et al., 2005).
As no sufficient calibration and orientation information was avai
lable, the whole block first needed to be adjusted. The method
for bundle block adjustment, including self-calibration of multi
ple devices and employing scene constraints to enhance the scene
geometry was introduced and tested in (Gerke and Nyaruhuma,
2009). The dense matching algorithm then was applied to sev
eral combinations of stereo images, and results were evaluated
through LIDAR data which was acquired from the FLI-MAP sys
tem.
4 http://www.flimap.nl
