International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
Figure 1.2 — 3D presentation. 
In order to improve the robust method, we are using orthogonal 
polynomials which permit the usage of interpolation functions 
of the polynomial kind, with no restriction on the degree of the 
polynomial. 
Moreover, in this paper, we also present an automatic 
method for DTM extraction which depicts the net roads, and 
using it as a first approximation of the DTM, then by iterative 
processing the true DTM is computed. To detect roads we use 
segmentation based on Normal directions, edge detections and 
height difference. 
For improving the extraction process, the Roads method and the 
Robust method have been merged. The results show success, at 
least in the qualitative aspect. 
2. Generating the DTM using the roads net 
2.1 The roads method outline 
We assume, prior to generating the DTM extraction, that 
roads patterns exist in urban areas. The basic idea is that we 
regard roads network as seed points in for determining the 
initial approximation of the DTM. To detect roads 
automatically we segment the data using Normal directions and 
height difference, these segments have been classified and the 
roads have been extracted. To execute this process we need to 
define the road accurately using its geometric properties. This 
explicitly defines the DTM borders. It is furthermore compared 
with the original DSM and if the difference falls below a 
predefined threshold (for example 0.3 m) the original DSM 
points are selected and included in the new DTM calculations. 
This process is repeated until the numbers of the added points 
in the last iteration falls blow a threshold. 
The LIDAR data is provided as points arranged by number of 
strips, these strips need to be adjusted to eliminate differences 
in the overlapping area, these differences caused mainly by 
changing the flight direction from one strip to another. 
The process of segmentation uses the original data after 
adjustment the strips. In the first step we calculate TIN model 
for the surface using Delauny algorithm, we also convert the 
data to a regular grid form to illustrate image processing tools 
and also to make the presentation of the results more easily. 
Figure 2.1 - First pulse 
  
Laser pulses can easily hit more than one object, especially 
when they hit trees. Using the first pulse (figure 2.1) can lead to 
problems because there are not enough points reflected from the 
ground and because trees may cover parts of the roads and 
make them thinner or discontinuous. In order to detect roads by 
Normal direction we use the last pulse (figure 2.2) assuming 
that it was reflected from the ground. Moreover to avoid 
discontinuous roads caused by traffic mass, high grid resolution 
is needed, which lead to long computation times and 
consumption of high memory storage. The size of the grid is set 
to 1 meter (figure 2.2). 
2.2 Segmentation and classification 
The results of the segmentation by Normal directions and 
height difference are shown in figure 2.3. In this segmentation 
we use Ah-0.5 m (difference in heights) and An=6 degrees 
(different in Normal direction). Road segments are considerably 
larger than any other segment, and also their area to boundary 
ratio approaches zero. The segments have been classified and 
the roads have been detected (figure 2.4). 
Internati 
  
2.3 Fina 
Using th 
the first 
DTM wi 
than 0.3 
to the D 
processir 
present 1 
(see are: 
processir 
some ca: 
seeds po 
points gr. 
Total nur 
Number « 
Number 
First iter: 
Second 11 
Third ites 
Total nur 
The final