CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
2.1.1 Pre-Processing
This step comprises the procedures from orientation of the input
data up to the DSM (Digital Surface Model) creation. For VHR
satellite imagery the orientation approach is based on the RFM
(Rational Function Model) (Vozikis et al., 2003). When dealing
with aerial imagery it is made use of GPS/INS information in
order perform direct georeferencing, and thus automated image
triangulation (Scholten and Gwinner, 2003). The DSM
extraction is performed by automated correlation procedures,
which nowadays are very mature and produce very good
results.
2.1.2 nDSM Creation
The goal is to derive the DTM (Digital Terrain Model) from the
DSM and subtract it from the DSM in order to produce the so-
called nDSM (normalized Digital Surface Model). This way all
extruding objects in the data set (including buildings) stand on
elevation height 0 (Figure 2).
Figure 2: DSM, DTM and nDSM.
2.1.3 Building Detection (Seeding)
This crucial step deals with the identification of potential
building candidates in the data sets (=determination of seed
points inside buildings). It is proposed to perform 2 statistical
analyses. First, perform a thresholding in the nDSM and filter
out all objects that are not taller than a certain height, and
second, perform texture analysis in the image data to keep only
roof-similar regions in the data set (Vozikis, 2004).
Figure 3: Computation of seed points (red asterisks) inside
potential building candidates by height-thresholding
and texture filtering.
2.1.4 Building Extraction
By applying the Hough Transformation (to an image of gradient
or of contours) the geometric properties of the buildings
(building edges and comers) are extracted. Our approach is
based on a stepwise, iterative Hough Transformation in
combination with an adaptive region growing algorithm
(Vozikis and Jansa, 2008). The general idea is to transform the
information in the image (feature space) into a parameter space
and apply there an analysis. It is a technique for isolating
features that share common characteristics. The classical Hough
transformation is used to detect lines, circles, ellipses etc.,
whereas the generalized form can be used to detect features that
cannot easily be described in an analytical way.
The mathematical analysis of the Hough Transformation is
described in detail in Gonzalez and Woods (1992).
Briefly it can be described as follows:
p = xcos(#)~ vsin(#) (1)
where p is the perpendicular distance of a line from the origin
and 0 the angle (in the range 0 to 7r) as illustrated in Figure 4.
To apply this function on the whole image, Equation 1 can be
extended as shown in Equation 2.
H(e,p)= \ jF(x,y)S(p- xcos(#)-ysin(#))
where 8 is the Dirac delta-function. Each point (x,y) in the
original image F(x,y) is transformed into a sinusoid p = xcos(0)
- ysin(0).
84
