Albert Baumgartner
“ribbon snakes" showed to be a very useful method to find a path between two road segments and to verify the con-
nection hypothesis. If the verification based on these criteria fails, an attempt is made to explain the gap between the
vo basic neighboring segments by information about the local context, e.g., due to a shadow cast by a building. These grouping
iles in a steps are applied iteratively, and the thresholds on the distance and the direction difference are relaxed step by step.
ules use Simultaneously with the relaxation of the thresholds short segments
odule II are removed. This elimination step is necessary because otherwise,
due to the relaxation of the grouping thresholds and due to the lim-
itations of the verification step, a lot of erroneous connection hy-
potheses would be accepted and would corrupt the further steps of
the road extraction. However, correct initial hypotheses are also
| model, removed.
d in the
ndidates After the generation of hypotheses for connections and their veri-
of edges fication, the road network is constructed (Fig. 4). Based on geo-
fference metric assumptions hypotheses for junctions are generated and ver-
ction of ified. Ideally, after this step all road hypotheses are connected, and
ion. there is a path between every pair of points on the extracted road
; à network. However, such a result cannot be expected, because the
quadr ie extraction is primarily based on local information and is reliable
adtilat- only in rural areas. In summary, module I uses only local infor-
d by the mation to establish connection hypotheses and to verify them. The
parts of network characteristics of roads are not optimally exploited. There-
fore, its most important feature compared to module II is the aspect
of local grouping. Apart from radiometric parameters which are
directly linked with the quality of the image, the threshold for the
elimination of the unconnected short segments is the most sensitive Figure 4: Results of local grouping.
parameter with respect to the quality of the results obtained with
this module.
3.2 Module II: Global Grouping
Module II is primarily based on the knowledge that roads have the function to connect different “important places,”
even if they are far away from each other. Roads form a (hierarchical) network that is mostly optimized to provide an
economic and convenient way for reaching different places. Because of this property, searching for the globally best
connection between such places is an essential step for road extraction. Moreover, since there usually exists only one
good connection between two “important places” (at least in open and rural terrain) the search can be restricted to the best
connection between two places.
This module starts the extraction of the road network with the ex-
traction of lines, calculates attributes for these lines and assesses
the probability of the extracted lines to be a part of a road network.
Based on local line attributes (e.g., straightness, length) which are
then compared to the knowledge about shape and reflectance prop-
erties described by the road model, each line gets a quality measure.
The endpoints of all lines are used as vertices of a graph. The lines
which connect the endpoints are edges in this graph and for all pairs
et more of vertices which are not connected by a line a quality measure for
gments the shortest connection is calculated. The quality measure of the
vis also “gap-edges” in this graph depends on purely geometric considera-
solution tions. The quality measures of lines and gaps are transformed by
linear fuzzy functions into values ranging from 0 to 1. An overall
fuzzy value of 1 means that the edge perfectly meets the properties
applied derived from the road model.
"creases
oreven Once the weighted graph is constructed, the next step is to select the
“important places”. Since this approach knows nothing about addi-
tional objects of the real world, e.g., buildings, industrial areas, or
eps. In other sites of interest, we define “important places” as lines that rep-
)etween resent portions of the road network with high probability. Hence, NT :
diomet- instead of connecting true “important places”, we try to connect Figure 5: Results of global grouping.
o-called pairs of high quality line segments (seed pairs). Additionally, the
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 61