gart,
orial
tion
gart,
ictie
Christian Wiedemann
AUTOMATIC COMPLETION AND EVALUATION OF ROAD NETWORKS
Christian WIEDEMANN, Heinrich EBNER
Chair for Photogrammetry and Remote Sensing
Technische Universität München, D-80290 Munich, Germany
Tel: +49-89-289-22572, Fax: +49-89-2809573
E-mail: wied|ebn@photo.verm.tu-muenchen.de
Working Group III/3
KEY WORDS: Networks, topology, evaluation, computer vision.
ABSTRACT
Road networks automatically extracted from digital imagery are in general incomplete and fragmented. Completeness
and topology of the extracted network can be improved by the use of the global network structure which is a result of the
function of roads as part of the transport network. This is especially — but not exclusively — important for the extraction
of roads from imagery with low resolution (e.g., ground pixel size > 1 m) because only little local evidence for roads can
be extracted from those images.
In this paper, an approach is described for the completion of incompletely extracted road networks. The completion is
done by generating link hypotheses between points on the network which are likely to be connected based on the network
characteristics. The proposed link hypotheses are verified based on the image data. A quantitative evaluation of the
achieved improvements is given.
New developments presented in this paper are the generation of link hypotheses between different connected components
of the extracted road network and the introduction of measures for the evaluation of the network topology and connectivity.
Results of the improved completion scheme are presented and evaluated based on the introduced measures. The results
show the feasibility of the presented completion approach as well as its limitations. Major advantages of the completion
of road networks are the improved network topology and connectivity of the extraction result. The new measures prove
to be very useful for the evaluation of network topology and connectivity.
1 INTRODUCTION
One of the major problems of image interpretation systems, in particular for road extraction, is that the results of the
extraction of primitives are incomplete (Steger et al., 1997). In order to overcome this problem, perceptual grouping
algorithms are used to extract meaningful entities from segmentation results. In systems for automatic road extraction,
hypotheses for the links between different parts of the segmentation result are often generated based on geometric criteria
like proximity and collinearity (cf., e.g., (Vasudevan et al., 1988, Ton et al., 1989, Mayer et al., 1997, Steger et al.,
1997, Baumgartner, 1998)). The hypotheses are then checked, e.g., based on the image data. Despite all these efforts,
until now, no fully automatic approach is able to extract road networks complete and correct from imagery. But some
approaches reach a completeness of 80% and more in open rural areas with about 95% of the extracted roads being correct
(Wiedemann et al., 1998).
Completeness and topology of the extracted network can be improved by the use of the global network structure which is
a result of the function of roads (Mayer, 1998). In this paper, an approach is presented for the completion of incompletely
extracted road networks. In general, grouping deals with the addition of new links as well as with the deletion of old
parts. The presented approach is only able to add new links. The approach is based on the function of roads as part of the
transport system.
An evaluation based on quantitative measures like completeness and correctness only, does not capture the improve-
ments of topology appropriately. Therefore, new measures for the evaluation are proposed, namely mean detour factor
and connectivity. These measures are calculated — similar to completeness and correctness — based on the com-
parison of the extraction result with reference data. They are intended to capture the topology of the extracted road
network.
In the following section a strategy for the generation of link hypotheses is proposed which makes use of the function of
roads. Section 3 deals with image based checking of the link hypotheses. In Sect. 4, new measures for the evaluation of
the achieved results are introduced. Results are presented, evaluated, and discussed in Sect. 5. The paper concludes with
a summary and an outlook.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 979