Internal
AUTOMATIC 3D BUILDING RECONSTRUCTION FROM AIRBORNE LASER m
SCANNING AND CADASTRAL DATA USING HOUGH TRANSFORM
2.1 He
Jens Overby ^ *, Lars Bodum“, Erik Kjems“, Peer M. Ilsge * Given €
line in :
* Centre for 3DGI, Aalborg University, Niels Jernes Vej 14, 9220 Aalborg, Denmark - (jeo, Ibo, kjems, ilsoe)@3dgi.dk pass thr
b for ve
KEY WORDS: Building, Reconstruction, Algorithms, Point Cloud, Geometry, Laser scanning, Virtual Reality, GIS points «
space v
the pars
3D city models are in the tele communication industry and for
analysing propagation of noise and air pollution in downtown
urban areas.
The so-called “position attributed information” is information
which is attached to a position, but isn't in itself geometrical. It
doesn't span any volume, neither has colours or surfaces — it can
be considered as zero dimensional. Most information in public
databases today is zero-dimensional, and again most of this is
attributed to a position. There is an increasing need for linking
3D geometrical and non-geometrical information, which gives
whole new possibilities for querying information in public
information servers. Queries are done directly by navigation in
3D. However, the roofs of buildings constitute a crucial part of
any city model. The whole of all building roofs generates a
“roof landscape”, which is by far the most dominating picture
when “flying” over a city model.
even more noise to the data. All these things considered, small
roof objects like chimneys, windows, and cornices become
noise instead of valuable data. It is also very difficult to identity
vertical gable ends separating different parts of a building roof.
In order to acquire a high density laser scanning, a helicopter
based system is necessary. Such systems can acquire point
densities of five to ten points per square meter [Baltsavias,
1999]. Algorithms dealing with high density laser scannings are
described in e.g. [Haala and Brenner, 1997] and [Vosselman,
1999]. Only very few areas are scanned with less than a 1xl
meter grid today, which is why there is a need for building
reconstruction algorithms, which can operate in low density
digital elevation models. This paper reports on our progress in
this particular field of research.
2. ROOF PLANE DETERMINATION
ABSTRACT:
A prob
Urban environments of modern cities are described digitally in large public databases and datasets of e.g. laser scanning and ortho approac
photos. These data sets are not necessarily linked to each other, except trough their geometry attributes (coordinates), which are around
mutually displaced and have a low degree of details. However, it is possible to create virtual 3D models of buildings, by processing line: x,
these data. Roof polygons are generated using airborne laser scanning of 1x1 meter grid and ground plans (footprints) extracted from
technical feature maps. An effective algorithm is used for fixing the mutual displacement of these datasets. The well known Hough The get
Transform is extended to 3D for extracting planes from the point cloud. Generally speaking, planes are rejected if the clusters of three d
points on these planes, do not span a considerable area. Furthermore, it is assumed that valid planes are close to parallel to one of the triplet (
ground plan lines. Points corresponding to each valid plane are subtracted from the original point cloud, and the Hough Transform is the tw
performed on the remainder. By this approach, the disturbing influence of already evaluated points, is avoided. Small variations of spheric:
the roof surface might lead to multiple slightly differing planes. Such planes are detected and merged. Intersecting planes are m (n.
identified, and a polygon mesh of the roof is constructed. Due to the low precision of the laser scanning, a rule-based post- the co
processing of the roof is applied before adding the walls. compor
Ej
| ny = CO:
1. INTRODUCTION mounted LIDAR system. However, the algorithms also work on
irregular grids. For building reconstruction this density is low, The dis
n Virtual 3D city models have become increasingly popular in considering that at least three points are required for defining a [Tis ex
j recent years. Often these models are used for supporting urban plane, and many of these points are inaccurate. Furthermore, the
planning processes and decision making. Other applications of regular grid is produced by interpolation, which has introduced s(0, @)
<
"x
Each sa
in paral
planes s
The cor
from su
One of the widely used methods for acquiring 3D geometrical A very important assumption is that building models can be sos
context is by using airborne laser scanners. The grid produced described by planar faces. Most algorithms for building I},je
by such scannings must be processed, ideally to reduce reconstruction work by detection of planar faces [Hoover et al., N,aret
information to the lowest possible level of entropy. This process 1996]. However, many of these algorithms require the ^
introduces removal of noise, intelligently adding missing computation of local surface normals. The problem is that these If s(0;
information, and removing the huge amount of redundant algorithms are extremely sensitive to noise [Vosselman, 1999]. value s,
information present in the surface scan, where in fact only 3 The plane detection can run into local minima, and thereby fail weight.
degrees of freedom (d.o.f) are necessary for defining any to find an obvious plane seen in the global perspective. by a nc
unbounded plane in 3D. (A normal vector where the length is describe
the distance from the origin of the coordinate system). A robust algorithm for this purpose is the Hough Transform
[Hough, 1962] extended to 3D [Schindler et al., 2003]. The The ‘res
The data used for the building reconstruction presented in this classical Hough Transform has been extensively used in the parame
article 1s a regular grid of 1x1 meter, acquired by an aeroplane
* Corresponding author.
in carte:
