nates. The
e building
ng and the
«tract form
case of the
f vertexes
rval of the
nately 2m.
h the exact
the case of
the rough
| due to the
can detect
the results
raight lines
ts the road
g edges are
nages. It is
f the high
correction
ecause the
ie segment.
zed Hough
ections and
3ased
th ISPRS
ing
II FIG
id
C
ion Of
FIG
1d
TOWARDS AUTOMATIC RELATIVE ORIENTATION FOR ARCHITECTURAL
PHOTOGRAMMETRY
Frank A. van den Heuvel
Delft University of Technology
Department of Geodesy
Thijsseweg 11, 2629JA Delft, The Netherlands
E-mail: F.A.vandenHeuvel@geo.tudelft.nl
Commission V, WG V/2
KEY WORDS: relative orientation, wide-baseline stereo, vanishing point detection, feature-based matching, architecture
ABSTRACT
A major challenge in close-range photogrammetry and computer vision is the automation of model acquisition from imagery.
Determining the relative position and orientation of close-range imagery is usually the first step in a procedure for modelling,
assuming that the camera has been calibrated beforehand. An essential part of an orientation procedure based on image content is the
establishment of correspondence between the images. The problem at hand is to find correspondence between two convergent
images and their relative orientation simultaneously and automatically. In computer vision this is called the wide-baseline stereo
problem. For the approach presented in this paper, a new feature-based matching procedure is designed that exploits the
characteristics of the application by applying generic knowledge of the construction of the building. The procedure relies on rigorous
statistical testing of constraints on the observations. The outline of the procedure is presented, as well as the results of experiments.
Relative orientation was successfully detected for two images with an angle of 65 degree between the optical axes. It is shown that
the procedure is robust with respect to unfavourable characteristics of the application, such as occlusions and repetitive structures in
the building facades.
1. INTRODUCTION
Determination of the relative orientation of images is a
prerequisite for object modelling. Although relative orientation
of two images includes relative position, it however excludes
the distance between the images and thus the scale of the model
remains undetermined. Camera calibration, i.e. the
determination of interior orientation (intrinsic) parameters, is
also essential, however, in photogrammetry cameras are usually
pre-calibrated. In this paper, the camera is assumed to be
calibrated, although imagery of unknown interior orientation
can be handled by applying a method for the estimation of
interior orientation parameters from vanishing points (van den
Heuvel, 1999).
Without using external measurement methods, such as the
Global Positioning System (GPS), relative orientation of two
images is based on corresponding features in the images. With
only two images, we have to rely on point features because
there is no information on relative orientation in corresponding
image lines. With at least seven corresponding points available,
the five parameters of relative orientation (in computer vision;
the parameters of the essential matrix) can be determined by a
direct solution (Fórstner, 2000). Automatic relative orientation,
therefore, is equivalent to solving the correspondence problem
automatically. This topic has been extensively studied in
photogrammetry, especially aerial photogrammetry (Heipke,
1997), and in computer vision (Pritchett and Zisserman, 1998),
(Matas et al., 2001).
Our goal is the automatic relative orientation of two widely
separated views, i.e. (strongly) convergent imagery. In
computer vision this problem is known as wide-baseline stereo.
Generic knowledge of the architectural application field is
applied in the method and this facilitates a robust solution of the
correspondence problem. For the method proposed here,
prerequisites for the image acquisition are limited to:
- No major image rotation around the optical axis (k less
than 45 degree).
- Image tilt (rotation around x-axis, œ) less than 45 degree.
- Overlap between the two images.
The automatic procedure consists of three main steps (Figure
1). In each step different types of a priori object knowledge, i.e.
knowledge on the construction of the building, are applied.
Firstly, straight lines are extracted from the two images. This
implies that the majority of building edges be assumed straight,
and that lens distortion is limited or removed in advanced by
resampling the image. Secondly, vanishing points are detected
for each image. The building is assumed to be built along three
orthogonal axes. The detection of at least two vanishing points
related to these object orientations, results in an ambiguous
orientation of the image relative to the building, and a set of
edges of which the object orientation is known. In the third and
last step, edges are intersected to points in image space, and
finally point correspondence and relative orientation are
detected simultaneously. The paper concentrates on this last
step that relies on the assumption of planar facades, in which
the object points recede.
The remainder of the paper is structured as follows. After
discussing related research in the next section, the developed
procedure for automatic relative orientation is described in
section 3. In section 4, results from several experiments are
analysed. Conclusions are drawn in section 5.
2. RELATED RESEARCH
In aerial photogrammetry the automatic relative orientation (as
a part of automatic triangulation) is commonly available in
digital photogrammetric workstations (Heipke, 1997). More and
more, external measurement devices — usually an integration of
GPS and INS — are used. For a range of aerial applications they
make triangulation even superfluous. External measurement of
image orientation is also used in close-range photogrammetry
(Teller, 2001). For finding correspondence from the image
itself, usually a feature-based matching technique is applied at
different levels of the image pyramid. Wang (Wang, 1998)
presents a structural matching method that requires an image
pyramid. The method is applied to aerial images, as well as to
terrestrial images of a building. In (Habib and Kelley, 2001) a
robust feature-based method is described that does not require
—227—
(BE POS ERS LE RRS
EEE EN
