SION
ng
1 approach to
are detected
Ie parametric
annels. Walls
plane patches
rformance of
nce, features
rtain. Height
he extraction
ies have led
ie data from
1ethods have
data (Ameri,
nsteiner and
methods has
' techniques.
) appeared in
i et al., 2000;
n, 2004). In
) used DSM
ted building
arried out on
i-source data
xplored. The
for fusion of
ed extraction
nbines aerial
] DTM from
S database in
ngs. In this
‚in the scene
and infrared
tion regions.
mation from
Is are finally
" planes.
an overview
nstruction of
3. Section 4
sing a plane
] results are
in 6.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
2. A FRAMEWORK FOR FUSION OF IMAGE DATA,
HEIGHT DATA AND 2D GROUND PLANS
The first step in a building extraction system is the detection of
buildings in the scene. In this work, buildings are detected by
projecting ground plans to image data. For this purpose, the
third dimension (height) of the ground plans is interpolated in
the height data.
Having buildings detected in the scene, the reconstruction part
is based on finding parametric forms of the model planar faces.
These planar faces are then intersected and. resulting plane
patches are assembled together to form a generic polyhedral
model. Roof planes are reconstructed by fusion of image and
height data through a split-and-merge process. This process
starts with segmenting the image data within the localized areas.
Height points from the DSM are projected to extracted image
regions and a robust regression method is employed to fit planar
faces to height points belonging to each image region. Regions
in which more than a single plane is detected are split and
neighbouring regions whose planes are coplanar are merged.
Vegetation regions are identified and discarded by computing
an NDVI measure derived from red and infrared channels.
Every planar face is attributed based on its slope and height
over the DTM. A planar face is attributed as non-roof if its
height over the DTM is smaller than a minimum tolerance;
‘otherwise it is attributed as flat roof if its largest slope is smaller
than a slope threshold or as slanted roof if the largest slope is
larger than the slope threshold.
Wall faces are obtained by reconstructing a vertical wall over
every line segment of the 2D ground plan. The average terrain
height, derived from DTM, defines the planar surface that lies
beneath the building. After the parametric forms of all planar
faces of the building are computed, every three planar faces are
intersected and the resulting vertex is verified to make sure it is
a correct model vertex. Verified vertices of each planar face are
sorted in order to form a planar patch. Planar patches form the.
final generic polyhedral model that can be visualized using a
graphical engine.
The basic assumption in this approach is that buildings are
formed by planar faces and that walls are vertical. In addition,
building roofs are assumed to be one of the following three
types: flat roof, gable roof and hipped roof. In other words this
approach aims for reconstructing simple building types using a
boundary-representation (B-Rep) modelling X scheme.
Nevertheless, more complex buildings such as buildings with
cross-gabled roofs can still be reconstructed by adopting a
Constructive Solid Geometry (CSG) modelling scheme. In this
way, similar to the method developed by Suveg and Vosselman
(2004), the 2D ground plan is first partitioned into rectangular
parts where each part is reconstructed using the plane patch
reconstruction method described above. These building parts are
then combined together to form the final generic model.
3. RECONSTRUCTION OF PARAMETRIC FORMS OF
THE MODEL PLANAR FACES
Buildings are localized in the image using ground plans and
height data. A split-and-merge process is applied to fuse image
and height data in the localized areas and derive the parametric
forms of roof planes. Walls are reconstructed by finding the
parametric forms of vertical planes built on the ground plan.
Iw following sections describe the above processes in more
etails.
3.1 Localization of buildings using ground plans and height
data
A 2D ground plan is usually stored as a polygon with an array
of corner points with X and Y coordinates in the world
coordinate system. The footprint of each building is localized in
the image by interpolating the height of every corner point of
the ground plan in DTM and projecting the resulting 3D corner
points to the image. Interpolation of heights in DSM with the
same procedure helps to find roof boundaries of the building in
the image assuming that walls are vertical and there is no eave
overshooting. Concatenation of these two polygons (footprint
and roof boundary) defines the actual area where the building
appears in the image.
3.2 Reconstruction of roof planes using image and height
data
Reconstruction of roof planes is based on image regions
extracted in areas where building candidates are detected.
Extraction of image regions is carried out using watershed
segmentation algorithm (Vincent and Soille, 1991). Extended
minima transform (Soille, 1999) is employed to control
excessive oversegmentation.
While a desirable segmentation is a partitioning of the image
into regions where each region corresponds to a single planar
face in object space, segmentation algorithms often result in
undergrown and/or overgrown regions. The purpose of the split-
and-merge process is to refine the result of initial segmentation
by making use of clues derived from the DSM. For this purpose,
height points are projected from DSM to extracted image
regions and a robust regression method is used to fit planar
faces to height points belonging to each image region. This
method is based on random selection of a finite set of samples
from data (trial estimates) (Fischler and Bolles, 1981). Least
median of squared residuals (Rousseeuw and Leroy, 1987) is
used to find the best sample and also outlier points. Each sample
contains three data points randomly selected from the DSM.
These points define a plane. For other points a residual value is
calculated as to how they fit into this plane. The sample with the
least median of squared residuals is selected for outlier
detection. Outliers are detected as points with residuals larger
. than a predefined tolerance and are treated as a new dataset to
determine whether they fit into a new plane. The plane fitting
process is iterated until no more planes can be fitted to data
points.
After planar faces are detected in each image region, the
segmented image is searched for regions in which more than
one plane is detected. Those regions are overgrown regions;
hence, they are split into two or more regions depending on the
number of detected planes. To detect and merge undergrown
regions, first a region adjacency graph is constructed by
tracking region boundaries in the segmented image. Plane
parameters of every two neighbouring regions enter a
coplanarity check and the two neighbouring regions are merged
if their associated planes are coplanar.
An example of the performance of the split-and-merge process
is demonstrates in figure 1. As can be seen in figure 1(B), the
initial segmentation results in an overgrown and an undergrown
region in the right part of the roof. The result of the split-and-
merge process is shown in figure 1(C) where the overgrown
region is split and two undergrown regions are merged to form a
correct roof region.