02-08-2011, 04:12 PM
ABSTRACT
This paper describes two developments in the automatic reconstruction of buildings from aerial images. The first is an
algorithm for automatically matching line segments over multiple images. The algorithm employs geometric constraints
based on the multi-view geometry together with photometric constraints derived from the line neighbourhood, and achieves
a performance of better than 95% correct matches over three views. The second development is a method for automatically
computing a piecewise planar reconstruction based on the matched lines. The novelty here is that a planar facet hypothesis
can be generated from a single 3D line, using an inter-image homography applied to the line neighbourhood. The algorithm
has successfully generated near complete roof reconstructions from multiple images. This work has been carried out as
part of the EC IMPACT project. A summary of the project is included.
1 INTRODUCTION
Reconstruction of buildings from aerial images has received
continual attention in the photogrammetry and computer
vision literature. One approach is to compute a dense
digital elevation model from multiple images using correlation
based stereo. However, the resulting dense depth map
is generally insufficiently accurate or complete to enable
the precise shape of buildings to be recovered (Berthod
et al., 1995, Baillard et al., 1998, Girard et al., 1998).
Thus most approaches have focused on reconstruction of
specific building models: rectilinear shapes (McGlone and
Shufelt, 1994, Roux and McKeown, 1994, Noronha and
Nevatia, 1997, Collins et al., 1998), flat roofs (Berthod et
al., 1995), or parametric models (Haala and Hahn, 1995,
Weidner and F¨orstner, 1995). Recently, more generic reconstruction
approaches involving multiple high-resolution
images have been proposed (Bignone et al., 1996, Moons
et al., 1998).
The difficulty of reconstruction in urban environments
comes from the complexity of the scene: the buildings are
dense and varied, and the resulting image boundaries often
have poor contrast. Consequently, feature detectors
fragment or miss boundary lines, and only an incomplete
3D wireframe can be obtained. This paper presents our
approach to solving this problem, which makes two contributions:
First, an algorithm for matching individual line segments
over multiple views; Second, an algorithm for computing
planar facets of the scene starting from the matched
lines. The key idea, which is common to both algorithms,
is that both geometric and photometric constraints should
contribute from all images.
In the line matching algorithm the match is computed using
multi-view geometry and photometric similarity measures
on the line neighbourhood in each image. Special attention
is paid to using multiple images to overcome the deficiencies
of the line segment feature extraction. In particular
fragmented lines are joined and missing edges recovered.
The algorithm is described in section 3.
In the plane generating algorithm 3D planar facets of the
scene are computed by using both lines and their image
neighbourhoods over multiple views. These surface facets
then enable both line grouping and, by plane intersection,
the creation of lines which were missed during feature detection.
The particular novelty of the method is in the
use of inter-image homographies (plane projective transformations)
to robustly estimate the planar facets. It requires
minimal image information since a plane is generated
from only a line correspondence and its image neighbourhood.
In particular two lines are not required to instantiate
a plane. These minimal requirements and avoidance
of specific building models facilitate the automatic reconstruction
of objects with quite subtle geometry located
within a complex environment. The algorithm is described
in section 4.
Our work is part of the European IMPACT project and we
begin with an overview of this project.
Download full report
http://googleurl?sa=t&source=web&cd=2&ve....1.30.9651%26rep%3Drep1%26type%3Dpdf&ei=WdQ3Tt6kCpHSrQf1qe3lDw&usg=AFQjCNEZ9fSstEE3LwHfx0o4Q1HdTB3cWA&sig2=7l4l6gvo3aTWhF241T4Pow