anbul 2004
OBJECT EXTRACTION FROM TERRESTRIAL LASER SCAN DATA FOR A DETAILED
BUILDING DESCRIPTION
Katharina Jülge } Claus Brenner
Institute of Cartography and Geoinformatics, University of Hannover, Germany
{Katharina.Juelge, Claus.Brenner} @ikg.uni-hannover.de
Working Group III/4
KEY WORDS: Feature, Extraction, Matching, Modelling. Close Range, Terrestrial, Laser scanning.
ABSTRACT
Building reconstruction has received increasing attention in the last few years. Many systems deal with reconstruction
from aerial laser scans and images. As there is an increasing demand for more detail in a building's description, terrestrial
data sources become more important. Preferably, extraction and reconstruction are realised as a fully automatic process.
This paper describes a new approach for segmenting and describing a building's facade. Basic models representing facade
parts such as windows are constructed as aggregations of geometric primitives like lines. Models are parameterized.
Edges are then extracted from laser scan data of a single building. Extracted edges are preprocessed using a length
filter. Subsequently, the previously defined models are fitted into the processed edge representation of the building using
a constrained search approach. The goal is to find multiple occurrence of a particular shape (i.e. multiple windows),
represented by an object model with a fixed parameter set, in one building. This approach works semi automatically with
a view to full automation in the future.
1 INTRODUCTION
This work is part of a research project which deals with
automatic derivation of 3D city models. Four data sources
are used for this:
|. airborne laser scans
t2
terrestrial laser scans
3. airborne images
4. terrestrial images
Objectives of the project include the fusion of multiple data
sets and automatic object extraction. This paper presents
one approach towards that goal. Images coming from one
of the named data sources, in this case range images de-
rived from terrestrial laser scans, are segmented. Edges
are extracted as low-level primitives. These edges are to
be aggregated into higher-level primitives; they represent
geometric structures present in the image. Because our
main concern are 3D city models, we have chosen build-
ing’s facades as an example. We are looking for geometric
structures which represent parts of the facades that occur
multiple times, such as windows, doors or ornaments. If
possible, these geometric structures are searched for in im-
ages coming from the other data sources in a way that cor-
relations are found.
In this paper, a procedure is developed to extract those
shapes. Model-based matching is used and constrained
search is applied to find successful matches.
* Corresponding author
1079
Our approach uses mostly 2D image processing tools, and
searching is done using 2D features, so preprocessing has
to be done to allow for distortion. To reduce the search
problem to two dimensions, features are extracted in the
facade using depth information and are then projected into
the fagade’s plane. Search is carried out matching the ge-
ometry of objects found in the façades.
2 RELATED WORK
Several papers deal with the extraction of façade structures.
For example, in (Wang et al., 2002) a system extracting
windows from the orthoimage of a fagade is described. A
consensus texture facade image is calculated from several
luminance-normalized facade images by weighted averag-
ing. The resulting image is deblurred and rectangles are fit-
ted into windows in the facade by an oriented region grow-
ing algorithm. This way, rectangles are iteratively fitted
into blobs representing windows so that they grow to as-
sume the window's size in the image. Extracted windows
are then grouped using clustering algorithms.
Another example is (Werner and Zisserman, 2002). Here,
the data source is an edge image with depth information
calculated over multiple views. Parameterized models are
fitted into facade structures such as windows or doors.
Models are three-dimensional and include structures com-
posed of straight lines such as boxes as well as more com-
plicated models which contain arcs. To select a particular
model, probabilities for models are determined using the
Bayesian rule over a set of training images.
In both examples, multiple photos are used as the data
source. In our case, we want to apply an algorithm falling
