ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002
polyhedral building models from LIDAR data only makes sense
if the point density is high enough so that a sufficient number of
data points is located at least in the most relevant planes of the
roofs. As the building outlines are difficult to be located
precisely, again ground plans are often used for that purpose.
Ground plans also reduce search space for the estimation of the
parameters of adjoining planar patches because the gradient
direction of such planes is usually perpendicular to the adjacent
polygon segment in the ground plan (Haala et al., 1998;
Brenner, 2000; Vosselman and Dijkman, 2001). Initial planar
patches are found by a segmentation of the DSM. Brenner
(2000) gives several methods for DSM segmentation, e.g., the
analysis of surface curvature, i.e., of changes in the surface
normal vectors, or a segmentation taking into account the
directions of the polygon segments of a ground plan.
As soon as the initial planar patches have been found,
neighboring patches are grouped (Baillard et al., 1999), and the
polygons delineating the borders of planar patches have to be
found. The latter task involves finding consistent intersections
at the building vertices (Moons et al., 1998). Finally, the 3D
border polygons have to be combined in order to obtain
consistent building models. At the building outlines, vertical
walls, and, finally, the floor have to be added to the model. A
coarse-to-fine strategy can be applied by first searching for the
most relevant structures in the data and using refined methods
for modeling the buildings in regions not being "explained"
sufficiently by the initial models (Vosselman and Dijkman,
2001). The problem of precisely determining the building
outlines has been tackled by Weidner (1997) by applying the
minimum description length principle for deciding on
regularizations.
2. WORK FLOW FOR BUILDING EXTRACTION
The work flow for the extraction of buildings from LIDAR
points is presented in figure 1. The first step is the interpolation
of a DSM and a DTM from the original data at an appropriate
resolution. Our method for DTM generation which performs a
classification of the original points into terrain versus off-terrain
points by robust estimation will be explained in section 3. From
this instance onwards, the models created by interpolation are
used, no longer the original data points.
By subtracting the DTM from the DSM and by applying a
threshold to the height differences, an initial building mask is
created which still contains vegetation and other objects. Binary
morphological operators and an analysis of the DSM texture,
i.e., of the local variations of the DSM normal vectors, are used
to eliminate these areas. The final results of building detection,
i.e., the individual building regions, are found by a connected
component analysis. Building detection by comparing the DTM
and the DSM is described in section 4.
In the building candidate regions, a plane segmentation based
on an analysis of the variations of the DSM normal vectors is
applied to find planar patches. These patches are expanded by
region growing algorithms. In the current version, the
neighborhood relations of these patches are determined, and a
simple model resembling the roof structure of the building is
created. In the future, neighboring planes will be grouped
consistently before the initial building models containing the
most relevant roof structures are created. In a post-processing
phase, the model has to be refined in order to contain details
originally not detected. The current state of our technique for
A - 296
geometrical reconstruction of roof structures from a DSM is
described in section 5.
Z LIDAR points / d
DSM segmentation
Region growing
Interpolation
Robust estimation
sa
/ ov // psu fo
a
Grouping of planes
Model generation
Initial building
models
Analysis of deviations
Model adaptation
Height thresholding
Initial building
regions
Morphological operators
Texture analysis
Building
regions
3D building
models
Figure 1. Work flow for building extraction from LIDAR data.
3. DTM GENERATION FROM LIDAR DATA IN
URBAN REGIONS
An algorithm for the automatic generation of DTMs in forested
regions from laser scanner data was developed at our institute.
This method is based on iterative robust interpolation of a DTM
grid, and it combines the elimination of off-terrain points and
the interpolation of the DTM grid in one process (Kraus and
Pfeifer, 1998). For the generation of a DTM in densely built-up
areas, this method has to be modified to work in a hierarchical
framework (Pfeifer et al, 2001). With this coarse-to-fine
approach it is possible to cope with relatively large areas
' without terrain points (e.g., large building blocks).
3.4 Robust Interpolation
In an iterative process the irregularly distributed LIDAR points
are weighted in a way that the modeled surface describes the
terrain. The classification of the points in terrain versus off-
terrain points is performed by thresholding the discrepancies to
the computed surface by user-specified tolerance values.
In a first step, a coarse approximation of the surface is
computed taking into account all available LIDAR points. Next,
the discrepancies, i.e., the differences of the heights of the
LIDAR points and the interpolated surface at the planimetric
positions of the LIDAR points, are computed. The discrepancies
are the parameters of a weight function which is used to assign
an individual weight to each point in the subsequent processes.
The interpolation of the DTM is repeated, the weights of the
LIDAR points being modulated depending on the discrepancies
of the most recent iteration. This iterative process is terminated
as soon as a stable situation or a maximum number of iterations
is reached. Two types of models are used in our algorithm, i.e.,
the functional model which defines the way the surface is
