In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Voi. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
229
GRAMMAR SUPPORTED FACADE RECONSTRUCTION FROM MOBILE LIDAR
MAPPING
Susanne Becker, Norbert Haala
Institute for Photogrammetry, University of Stuttgart
Geschwister-Scholl-Straße 24D, D-70174 Stuttgart
forename.lastname@ifp.uni-stuttgart.de
KEY WORDS: Architecture, Point Cloud, Urban, LiDAR, Facade Interpretation
ABSTRACT:
The paper describes an approach for the quality dependent reconstruction of building facades using 3D point clouds from mobile
terrestrial laser scanning and coarse building models. Due to changing viewing conditions such measurements frequently suffer from
different point densities at the respective building facades. In order to support the automatic generation of facade structure in regions
where no or only limited LiDAR measurements are available, a quality dependent processing is implemented. For this purpose,
facades are reconstructed at areas of sufficient LiDAR point densities in a first processing step. Based on this reconstruction, rules
are derived automatically, which together with the respective facade elements constitute a so-called facade grammar. This grammar
holds all the information that is necessary to reconstruct facades in the style of the given building. Thus, it can be used as knowledge
base in order to improve and complete facade reconstructions at areas of limited sensor data. Even for parts where no LiDAR
measurements are available at all synthetic facade structures can be hypothesized providing detailed building geometry.
1. INTRODUCTION
Due to the growing need for visualization and modelling of 3D
urban landscapes numerous tools for the area covering
production of virtual city models were made available, which
are usually based on 3D measurements from airborne stereo
imagery or LiDAR. This airborne data collection, which mainly
provides the outline and roof shape of buildings, is frequently
complemented by terrestrial laser scanning (TLS). However, the
applicability of standard TLS is usually limited to the 3D data
capturing of smaller scenes from a limited number of static
viewpoints. In contrast, the application of dynamic TLS from
moving platforms allows the complete coverage of spatially
complex urban environments from multiple viewpoints. One
example of such a mobile mapping system, which combines
terrestrial laser scanners with suitable sensors for direct
georeferencing, is the StreetMapper system (Kremer and
Hunter, 2007). This system enables the rapid and area covering
measurement of dense 3D point clouds by integrating four 2D
laser scanners with a high performance GNSS/inertial
navigation system. By these means accuracy levels better than
30mm have been demonstrated for point measurement in urban
areas (Haala et al., 2008).
In general, such systems allow for an efficient measurement of
larger street sections including the facades of the neighbouring
buildings. However, depending on the look angle during the
scanning process, strong variations of the available point
densities at the building facades can occur. Such viewpoint
limitations and occlusions will subject the collected point cloud
to significant changes of accuracy, coverage and amount of
detail. For this reason, the following interpretation of the
measured point clouds will be hampered by considerable
changes in data quality. Thus, algorithms for automatic facade
reconstruction have to be robust against potentially incomplete
data sets of heterogeneous quality. For this purpose, dense point
cloud measurements for facades with good visibility are used in
our approach to extract rules on dominant or repetitive features
as well as regularities. These rules then are used as knowledge
base to generate facade structure for parts or buildings where no
sensor data is available. By these means bottom-up and top-
down propagation of knowledge can be combined in order to
profit from both reconstruction techniques. The production
rules, which are automatically inferred from well observed and
modelled facades, are represented by a formal grammar.
Such formal grammars are frequently used within knowledge
based object reconstruction to ensure the plausibility and the
topological correctness of the reconstructed object elements.
Lindenmayer-systems (L-systems), which can be applied to
model the growth processes of plants, are well known examples
of formal grammars (Prusinkiewicz and Lindenmayer, 1990).
So-called split grammars are introduced by Wonka et al. (2003)
to automatically generate architectural structures from a
database of rules and attributes. Similarly, Müller et al. (2006)
present a procedural modelling approach for the generation of
detailed building architecture in a predefined style. However,
the variety of facade structures which can be generated is
restricted to the knowledge base inherent in the grammar rules
or model libraries. Thus, the appearance of facade elements is
limited to prespecified types, even when leaving some freedom
in the values of their parameters. Another problem while
applying such approaches for object reconstruction is that
manual interaction is required to constitute suitable building
styles and translate them into some kind of model or grammar
description. For this reason, several approaches aim at deriving
such kind of knowledge from observed or given data. For
example, Ripperda (2008) derives prior facade information
from a set of facade images in order to support the stochastic
modelling process. However, existing methods which try to
derive procedural rules from given images as proposed by
Müller et al. (2007) or Van Gool et al. (2007) still resort to
semi-automatic methods. The same holds true for the work of
Aliaga et al. (2007). They present an interactive system for both
the creation of new buildings in the style of others and the
modification of existing buildings. At first, the user manually
subdivides a building into its basic external features. This
segmentation is then employed to automatically infer a grammar