it 
at 
  
    
Interna 
  
this case the terrain objects — can be excluded after subsequent 
classification process. 
Favourably, the segmentation our relevant 3D objects is carried 
out in such a normalised surface model (nDSM) by a special 
region growing algorithm which extracts and separates 3D 
object areas. Starting point (crystallisation point) is a user- 
defined neighbourhood of a point (e.g. N8) in this data set 
where all points exceed the minimal object height above ground 
(e.g. 2.0m). During an iterative process all new neighbouring 
points are included in this segment which have a height 
difference smaller than the maximal acceptable one 
(homogeneity criterion). This procedure results in separated 
areas of 3D object while very small and low objects are 
excluded. Fig. 3 and 4 shows the segmented objects of test site 
Salem and Karlsruhe. 
3.3 Feature extraction 
Inside the segmented object areas specific features for 
distinction ‘of the relevant classes buildings, vegetation and 
terrain are extracted: 
e Gradients on segment borders 
eo Height texture 
e  First/last pulse differences 
e Shape and size 
e Laser pulse intensities 
The formerly tested feature direction of normal vectors was 
excluded, because of ambiguous results at smaller objects. 
Significant gradients along the border of segments contribute 
mainly to a discrimination of buildings/vegetation on one hand 
and terrain objects on the other hand. While buildings and trees 
generally show a high amount of border gradients in 
laserscanning data (70% - 100%) most segmented terrain 
objects — even if sharp relief edges are included — have at least 
at some parts of the segment borders smooth transitions to the 
surrounding terrain model. Therefore, the amount of significant 
border gradients decreases below 50% in these cases. 
  
Figure 5. Grey coded height texture (Laplace operator) 
In contrast height texture and first/last pulse differences allow 
the distinction of vegetation and buildings. Taking the shape of 
building roofs into account exclusively those height texture 
| 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 | 
| 
| 
| 
  
  
parameters seem to be useful that model the deviations from pese 
oblique planes which fits very well to the characteristics of ees 
buildings in laserscanning data. Suitable results can be obtained gn 
by the Laplace operator (Maas, 1999) or by /ocal curvature | mm 7 
(Steinle, Voegtle, 2001), i.e. the difference of subsequent | Ke 
gradients in the four directions across a raster point. Inside the respect 
roof planes of buildings small height texture values will be | m 
obtained while vegetation objects causes significant higher 
values (Fig. 5.) The differences of first and last pulse | 
measurements show a similar characteristic. Building roofs | 
normally consist of solid material, so - dependent on the slope | 
of the roof plane - no or only smaller differences between first | 
and last pulse measurements can be observed. In contrast at | 
vegetation objects with its canopy partly penetrable for laser | 
beams larger differences will occur. Additionally, a new | 
parameter was developed and tested, the local variance | 
differences of first and last pulse measurements. But this | 
parameter contains nearly the same information as height | 
texture and first/last pulse differences and, therefore, was | 
excluded for the further investigations. On principle high | 
texture values as well as high first/last pulse differences can be 
observed at the border of both, buildings. and vegetation. | 
Therefore, only the interior part of the segment areas can be | 
used for determination of these parameters to avoid | 
disturbances by this effect. 
Fig 
| 
A 
“10 
Figure 6. Grey coded first/last pulse differences Fe 
The shape of segmented object areas may contribute to the E er 
discrimination of artificial (man-made) objects (e.g. buildings, infoni 
bridges etc.) and natural ones (e.g. trees, groups of trees, rough of 1 
terrain or combination of both). For determination of shape retlect 
parameters the contour lines of each segment has to be ied i 
extracted. Because working with segments of uniform (pixel) Sdn 
values and clearly defined borders a simple edge tracking tens 
algorithm can be applied to provide the 2D contour lines. After | in Fig, 
smoothing these lines, e.g. by the well-known Douglas-Peucker Some 
method (Douglas & Peucker 1973), shape and size of these RMS 
polygons can be analysed. Former investigations have shown set 
that commonly used standard parameters like roundness, wil 
compactness etc. don't fulfil the requirements which are 
necessary to distinguish between the object shapes in this Fuzz 
application. Therefore, alternative parameters had been ; 
developed like geometry of the n longest lines, where at first the The s 
n longest lines of a contour polygon are selected (e.g. n=4). precec 
These lines are analysed in terms of parallelism and object 
orthogonality. A measure is calculated which is 100 for perfect 
410