- Approximate localisation of the building is assumed
known.
- The building must be oriented in three main
directions; one vertical, and two horizontal and
perpendicular.
The last requirement, does not imply that the building
must consist exclusively of boundaries in the three main
directions, only that such boundaries must exist. In the
current state, parallel, horizontal lines may only be
connected by lines perpendicular to them. In effect, this
means that 3D rectangles (tilted as well as horizontal) can
be found. General 3D parallelograms can not be found,
since neither of two opposite lines is perpendicular to the
other two.
The main parts of the system are shown in Fig. 1.
The implementation of the described approach is not done
as a streamlined production tool, but more as a loosely
connected system of related programs. For this reason, no
calculation times or efficiency numbers are presented.
The main image feature used for the
correspondence task are straight lines. Straight lines with
high precision can be found by standard methods. In
general, these methods give the end points of an isolated
line. If regional descriptors, like average grey level, or the
topology of the lines are desired, a region segmentation
must be performed. We believe that both straight lines
and regional information are needed, and use a region
segmentation, that uses straight lined boundaries of the
regions (Wiman 1995).
The vast majority of all buildings fulfil the criteria
that some lines are horizontal, and that these lines are
oriented in either of two perpendicular angles when
projected to the XY plane. Most additions to buildings,
that may be added to a geographic data base in an
automated map revision process, obey these rules as well.
The lines extracted by the region segmentation are
therefor first examined with respect to object space
orientation. The two main, horizontal and perpendicular
directions are determined by the examination.
The lines that have contributed to the definition of
the main directions are then analysed in a clustering
algorithm (Axelsson 1994), which accumulates
intersections of these lines. Once again, evidence from
each image is accumulated in object space and then
analysed in object space. The major clusters, which have
contributions from all images, form in principal endless
horizontal lines in either of two perpendicular
orientations. These endless lines are currently truncated
based on expected size of the object.
Each pair of parallel lines forms a plane, which is
hypothesised as an object plane. Each 2D line that is
inside the projected window of a hypothesised plane is
analysed whether it (7) fits to the plane and (ii) intersects
the parallel 3D lines in right angles. If so, their
intersection points on the 3D lines are computed. For true
vertical structures, the intersection points will be
approximately the same for a large number of 2D lines,
which thus form a cluster. False planes do not have any
pronounced clusters. The outlines of a true plane is
determined by finding the strongest two clusters. This is
the third time an image-to-object accumulation followed
954
by an object space analysis is performed. In separating
false plane from true and defining the outlines of the true
planes, radiometric evidence, collected from the original
segmentation, is used in combination with the geometric
evidence.
Segmentation
Region segmented | Object related using
image with straight | straight lines,
Image | space boundaries evaluated by the
bject| space MDL principle
Main Directions
Find the main Assumes the object
directions of the has two main
object directions that are
horizontal
Find the Use image relations
intersections of the | and object know-
image lines of the ledge,
main directions evaluated by the
MDL principle
Object Planes
Find the 3D planes | Use geometrical
delineating the constraints and
object radiometry of the
regions to find the
delineation’s of the
3D planes
Present the most
probable object
model
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Figure 1. The main parts of a system for autonomous
description of 3D structures.
3.2 Initial data
Most systems for automated description of buildings use
either one image, e.g. (Braun 1994, McGlone et.al. 1994,
Lin et.al. 1995) or two images, e.g. (Jamet et. al. 1995,
Roux et.al. 1994). Our system is specifically designed to
handle more than two images without prohibitive increase
in search space.
The approximate localisation, but not orientation or
shape, of the building is assumed known. The input thus
consists of digital image patches, one from each aerial
photograph in which the building is imaged. The interior
orientation of the camera(s) and the exterior orientation of
the images must be known.
We will illustrate our strategy with an example, using
a building with a simple geometric shape. The building
was imaged in six aerial photographs. One of the six
