3D building reconstruction from airborne lidar point clouds fused with aerial imagery - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 71

1995). Building boundaries, the intersection of the buildings
with its surroundings (e.g., ground), need to be derived from
the classified building points. Typically, the building boundary
generation is initiated by detecting a coarse approximation of
the outline, followed by a generalization and a regularization
(Sampath and Shan, 2007; Jwa et al ., 2008). Two fundamen-
tally different approaches for building reconstruction can be
distinguished: model-driven and data-driven approaches.
In the model-driven methods a predefined catalog of roof
forms is prescribed (e.g., flat roof, gable roof, etc.). The shape
and position parameters are determined by fitting models to
lidar point clouds (Weidner and Forstner, 1995; Haala, Brenner
and Anders, 1998; Maas and Vosselman, 1999). These methods
will lead to a reliable reconstruction if all the constraints in
building models are well satisfied and produce pretty results.
For instance, an algorithm may assume that there exists a main
orientation of the building and all edges are either parallel or
perpendicular to that orientation. This is especially appropriate
for low point densities. An advantage is that the final roof shape
is always topologically correct. A disadvantage is, however, that
complex roof shapes cannot be reconstructed, because they are
not included in the catalog.
The data-driven methods do not assume specific building
shapes for a scene, only making a few realistic assumptions such
as the vertical wall constraint. A building can be expressed by
bounding surfaces that may be described as planar surfaces or
triangles. Although some algorithms can produce good building
shapes, they often use complex plane detection techniques such
as clustering of triangles based on triangulated irregular networks
(TINs), 3D Hough transformation and clustering of 3D points,
which often result in a heavy computational burden (Maas
and Vosselman, 1999; Wang and Schenk, 2000; Vosselman and
Dijkman, 2001). Some algorithms produced polyhedral models
with low quality in shape (Gamba and Houshmand, 2002).
In the data-driven methods the roof is ''reassembled'' from
roof parts found by segmentation algorithms. The results of
the segmentation process are sets of points, each one ideally
describing exactly one roof face. Some roof elements (e.g., small
dormers, chimneys, etc.) may not be represented. The challenge
is to identify neighboring segments and the starting and ending
point of their intersection.
Building reconstruction is based on the assumption that indi-
vidual buildings can be modeled properly by a composition of a
set of planar surfaces. Hence, it is based on a reliable 3D segmen-
tation algorithm, detecting planar faces in a point cloud. This
segmentation is of crucial importance for the outline detection
and for the modeling approach (Dorninger and Pfeifer, 2008).
Segmentation allows for a decomposition of a building in a lidar
point cloud into planar surfaces and other objects. This requires
the definition of a homogeneity criterion according to which
similar items (e.g., points) are grouped. As homogeneity criteria,
approximate height similarity or/and approximate normal vector
similarity are commonly used. Determination of planar faces for
roof modeling from point clouds acquired from airborne plat-
forms is studied in the literature (Maas and Vosselman, 1999;
Lee and Schenk, 2002; Rottensteiner et al ., 2005). To reduce the
complexity of the problemand to increase the performance of the
implementation, again, 2.5D grid representations are commonly

used instead of the original points. This requires the definition
of a reference direction (e.g., the vertical z -axis) to resample the
given points to a regular grid defined as a scalar function over
the horizontal xy -plane. Thus, only one distinct height value can
be assigned to an arbitrary pair of xy -coordinates. Advantages
of2.5Dapproachesareapossiblereductionoftheamountof
input data and the implicitly defined neighborhood by means of
the grid representation. By contrast, for processing original point
clouds, such a neighborhood (e.g., for the estimation of normal
vectors) has to be defined explicitly (e.g., Filin and Pfeifer, 2006).
Unfortunately, the grid resampling process introduces smooth-
ing effects especially at sharp surface structures. Segmentation
approaches based on range images suffer from these restrictions
as well (e.g., Hug and Wehr, 1997; Maas and Vosselman, 1999).

6.2 Our building
reconstructionmethod

In recent years, lidar data has been widely applied in urban 3D
building extraction. A variety of methods have been proposed
for this purpose. However, difficulties in building detection and
reconstruction exist when lidar point cloud data is used alone
due to its irregular distribution, lacks of spectral, texture and
shape information. Also, edges from lidar point cloud data are
jaggy due to the relatively low sample rate. Aerial images provide
detailed texture and color information in high-resolution, mak-
ing them necessary for texture data and appealing for extracting
detailed model features. However, reconstruction from stereo
aerial images only leads to sparse points, which makes them
unsuitable for reconstruction of complex surfaces, such as curved
surfaces and roofs with slopes. As such the building extraction
based on either single image data or single lidar data cannot
reach a satisfying result. To this end, this chapter presents a
combination of lidar point cloud data and color aerial image data
for 3D building extraction.

6.2.1 Our strategy using fused data

The proposed building extraction strategy is based on the sequen-
tial determination of individual building models from both lidar
point cloud data and aerial image data. Figure 6.1 shows the
workflow of our strategy.
Both an unstructured lidar point cloud with the first and last
returns and a color aerial image covering the same area are used
as the input. Spatial registration of lidar point cloud data and
optical image is performed as data preprocessing.
At the building detection step, filtering is first used to cate-
gorize the lidar data into two classes: on-terrain and off-terrain
points. The filtered lidar points are then classified into three
feature classes, named buildings, trees, and ground (or on-
terrain points). The line features extracted from the aerial image,
geometric information, such as height information, discretemea-
surement, and the height difference between the first and last