International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
Figure 3. Parcel extraction example with starting polygon (black) 
and extraction result without smoothing (red) 
  
Figure 4. Extracted parcel after smoothing 
4.3 Line Extraction 
The tool for the measurement of linear features (in particular 
road networks) is based on a line tracking algorithm, where the 
user first defines a starting point and the measurement direction 
in the image. After that the procedure starts and automatically 
measures points along the middle axis of the particular line. Ex- 
isting lines and crossings are snapped and nodes are generated, 
leading to a topologically connected network, which is very im- 
portant in the case of a road network. The lines are smoothed 
and the average width of the line segments is computed auto- 
matically, too. Finally, the resulting topologically connected 
road network is smoothed and the operator can key in the miss- 
ing GIS attribute values or select them from a domain table. 
In principle, approaches for the semi-automatic extraction of 
linear features can be classified into two categories: 
|. Line trackers i.e. approaches that follow a linear feature 
that is specified by some initial information (see e.g. Groch 
1982, McKeown & Denlinger 1988, Vosselman & de 
Knecht 1995). 
2. Approaches that determine a linear feature that connects at 
least two given points (see e.g. Fischler et al. 1981, Grün & 
Li 1997): 
We decided to implement an approach of the first category, 
because the handling of such an approach seems to be more 
intuitive, especially for users who are normally digitising linear 
features point by point. The main principle of our line tracking 
algorithm corresponds to the one presented in Vosselman & de 
Knecht (1995). 
From the user's point of view, the extraction process consists of 
five major parts: 
1. Interactive setting of the initial information for the line 
tracker. 
2. Automatic tracking of the linear feature. 
3. Interaction, if necessary. 
4. Automatic determination of the line width. 
5. Line smoothing. 
The initial information that must be available for the algorithm 
consists of the starting point, the starting direction, and the 
width of the linear feature. To this end, the starting point must 
be measured and a point in the direction of the linear feature 
must be chosen. In order to reduce the number of necessary user 
interactions in the initialisation phase, the width of the linear 
feature is determined automatically and visualized for control 
purposes. Here the user has the possibility to correct the 
automatically determined line width, if necessary. After that, the 
line tracking can be started immediately. 
The automatic line tracking process consists of the following 
steps (see Figure 5): First, a grey value profile is determined 
perpendicular to the line direction. This grey value profile serves 
as reference for the following tracking process. Then the 
position of the next point along the axis of the linear feature is 
determined by a fusion of the predicted line position and the 
measured line position. The prediction is based on the current 
direction of the line and the measurement is based on a 
correlation of the reference profile with a search profile that is 
extracted perpendicular to the line direction at the predicted 
position. Subsequently the approach decides whether the new 
point can be accepted or the line tracking must be stopped e.g. 
due to low correlation values at the last positions. 
The algorithm follows the line as long as possible. It stops in 
one of the following cases: 
e It reaches an already extracted line or junction. 
e It is terminated interactively by the user. 
e The line cannot be found anymore. 
If an already extracted line or junction has teen reached, it is 
snapped and the generation of a (new) junction is proposed. 
This junction can be accepted by the user, it can be 
geometrically shifted or it can be rejected. If the line tracking has 
been stopped by the user or because the line could not be found 
anymore, it is possible to correct the current extraction result 
interactively e.g. by deleting the last few extracted points. 
Afterwards, a new starting direction and, if necessary, a new 
  
Th 
ass 
par 
Bef 
nur 
Raı