International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
The disadvantage of the known model-driven approach is in the
fact that it is, in comparison to others, inflexible regarding the
building’s ground plan. Many segmentation-based approaches
are based on raster data, enabling the use of standard image
processing software tools. Region-growing based approaches
have the advantage that the detected roof planes fit very well
into the laser scanner points. They may however be sensitive to
errors or gaps in the data
The additional use of ground plan information has the
advantage that the position and extension of a building are
known. This replaces error-prove segmentation methods, and it
simplifies the modelling of a building substantially.
Nevertheless, the method presented here is a generic method
based purely on laserscanner data, exploiting and proving the
potential of this data.
1.2 Goal and idea
The goal of the here described method is the automatic
reconstruction of building models from raw airborne laser
scanner data without any additional information. In addition to
a virtual 3D model of the reconstructed building, the ground
plan of the building will be generated. The chosen procedure is
a coarse-to-fine method, generating a rough building model in a
first stage and then successively refining it by an analysis of the
point cloud.
The basic idea of the chosen procedure is a simulation of the
behaviour of a human operator when analysing a point cloud in
order to recognise a building. This procedure is characterised
by the rotation of a point cloud into specific projections. Most
information is received by the view of the gable of a building.
In Figure 1-2 it is demonstrated that roof faces are reduced to
lines if projected onto a plane parallel to the gable.
From this line of sight the roof type as well as the number, the
inclination and the width of roof areas can be recognized. This
information is used in the method to detect roof faces. For this
purpose, the orientation of the building is determined and the
data points are projected onto the appropriate vertical plane. In
these projections, lines are searched, which represent roof areas.
These lines are finally extended to roof areas and grouped to a
building model. Detecting roof planes is thus reduced to a two-
dimensional parameter space, what reduces the complexity and
ensures a more efficient performance of the algorithm.
z
A "s
a .
. s vs
tan ta to
» zy» > E 2242 "ei
en pad RSA LTR
oon, BIN NS 048 = re
...
TE . irri
. .*
Ce
s"* 5.
Figure 1-2: A building's point cloud projected onto a plane
parallel to the gable side of the building
2 DATASET
The described method was tested on three different data sets,
which differ by their point density and the kind of the buildings:
The first data set is an alpine area in Switzerland. The point
density amounts to approximate one point per square meter.
The area is characterised by single houses and simple roof
morphologies. It predominantly contains simple or combined
saddle roofs. A special peculiarity of the buildings in this data
set is that they are partly built into the slope and possess large
roof overhangs. The latter is to be taken into account for the
determination of ground plan information. 229 buildings of this
data set have been chosen to determine the effectiveness of the
developed modelling procedure.
The second data set includes the city of Freiberg which is on
rather flat terrain in Saxony/Germany. This test area is
characterised by a high density of buildings including various
large and complex buildings as well as different roof shapes.
The point density amounts to 3-5 points per square meter.
The accuracy of both data sets is estimated «30cm in position
and +20 cm in height. In both cases the raw irregularly
distributed laserscanner data were used. Segmented point
clouds of the data set contain an individual building, some
surrounding ground points as well as points of adjacent
vegetation. The point density of the data sets results from
several highly overlapping flight strips.
As a third data set an area within the city of Dresden
(Saxony/Germany) was used. This dataset contains rasterised
data with a point distance of one meter. It serves as a
comparison of the modelled buildings with data obtained by
classical terrestrial measurements.
3 METHOD
3.1 Model
The variety of the existing roof shapes places different
requirements on an algorithm for building model
reconstruction. Not all details and shapes, which arise in
reality, can be modelled. Therefore it is first necessary to
develop a certain model conception of a building. For this
purpose initial considerations have to be made, which roof
shapes arise most frequently and which common characteristics
arc valid for the majority of the buildings.
It is assumed that buildings consist of plane surfaces. Thus, the
algorithm can be based on the search for straight lines in
2D-projections. The fundamental roof shapes, which should be
modelled with the available algorithm, are gable roofs, pent
roofs and hipped roofs as well as combinations of these basic
shapes. A first assumption is that buildings possess maximally
two main directions that are orthogonal to each other. The
normal vectors of all roof faces should be orthogonal to one of
the main directions. The main directions of the building are
defined by the orientation of the ridges and/or by the orientation
of the building edges. In order to model more complex
buildings, such as buildings with combined roof shapes, it is
necessary to make further assumptions. A building can possess
multiple ridges of different height. They are parallel or
orthogonal to each other. Besides, it is taken that the lower
eaves (the gutter) is parallel to the ridge and the ridge is parallel
to a horizontal plane. The walls of a building are plane vertical
surfaces and are attached to the eaves, under consideration of a
Intern
certail
a ridg
3.2
(
The.
point
The ir
means
a)
3.3
R
b
E
[
«us pe (x. PM mm LA eu
D
As the
projec
ridge €
determ
ground
histogr
within
separat
With t]
search
main |
Figure
The p
orienta
cloud.
points :
height
search
the ran
lines c
buildin
of vege
The mc
that is :
now ro
directio
the y-a>
