ISPRS Commission III, Vol.34, Part 3A , Photogrammetric Computer Vision“, Graz, 2002
URBAN ROAD NET EXTRACTION INTEGRATING INTERNAL EVALUATION MODELS
Stefan Hinz and Albert Baumgartner
Chair for Photogrammetry and Remote Sensing,
Technische Universität München, Arcisstrafe 21, 80290 München, Germany
{Stefan.Hinz, Albert.Baumgartner}@bv.tum.de — www.RemoteSensing-TUM.de
Commission III, Working Group II[/4, III/7
KEY WORDS: Road Extraction, Urban Areas, Quality Measures, Internal Evaluation, Self-Diagnosis, Image Understanding
ABSTRACT
This paper focuses on internal quality measures for automatic road extraction from aerial images taken over urban areas. The motivation
of this work is twofold: Firstly, any automatic system should provide the user with a small number of values indicating the reliability
of the obtained results. This is often referred to as "self-diagnosis" and is in particular a crucial part of automatic image understanding
systems. Secondly, and more important in the scope of our research, a system designed for the extraction of man-made objects in
complex environments (like roads in urban areas) inherently implies many decisions during the extraction process. Such decisions are
highly facilitated when both low level features and high level objects are attached with confidence values indicating their relevance for
further processing. The basic idea for defining evaluation criteria from which the confidence values can be calculated is to split the
components of a semantic object model into two different types. The model components of the first type are used for extracting features,
i.e., parts of the object, and the components of the other type serve as criteria for evaluating the quality of the extracted features. For
guaranteeing an unbiased evaluation one has to ensure that model components belonging to different types are independent from each
other (at least theoretically). We illustrate this concept by our system for road extraction in urban areas. Examples are given for both
low level features like lines and ribbons as well as higher level features like lanes and road segments.
1 INTRODUCTION
From a practical point of view, research on automatic road extrac-
tion in urban areas is mainly motivated by the importance of geo-
graphic information systems (GIS) and the need for data acquisi-
tion and update for GIS. This demand is strikingly documented in
the survey on 3D city models initiated by the European Organiza-
tion for Experimental Photogrammetric Research (OEEPE) a few
years ago (Fuchs et al., 1998). Applications of road extraction in
urban areas include analyses and simulations of traffic flow, esti-
mation of air and noise pollution, street maintenance, etc.
From the scientific perspective, the extraction of roads in com-
plex environments is one of the challenging issues in photogram-
metry and computer vision, since many tasks related to automatic
scene interpretation are involved. In order to cope with the high
complexity of urban scenes, our extraction system integrates de-
tailed knowledge about roads and their context using explicitly
formulated models. The road model includes, for instance, small
sub-structures such as markings but also knowledge about the
global network characteristics of roads, while the context model
describes relations between roads and other objects, e.g., build-
ings casting shadows on the road or cars occluding parts of a lane.
This makes it possible to extract roads even if their appearance is
heavily affected by other objects.
The work presented in this paper focuses on the development of
internal quality measures for automatic road extraction. The mo-
tivation of this specific aspect within an object extraction sys-
tem is twofold: Firstly, any automatic system should provide the
user with some values indicating the reliability of the obtained re-
sults. This is often referred to as "self-diagnosis" (Fórstner, 1996)
which is a crucial part of automatic image understanding systems,
in particular, when designed for practical applications. Secondly,
and more important in the scope of our research, confidence val-
ues also play an important role for the reliability of the extraction
itself, since they highly facilitate inevitable decisions which have
to be made during the extraction process. Consider, for instance,
competing road hypotheses extracted from multiple overlapping
images which must be combined into an unique road network.
The correct selection becomes much easier when the hypothe-
ses are attached with confidence values indicating their quality
and relevance for further processing. Such situations often occur
within image understanding systems designed for the extraction
of cartographic objects from natural scenes. Hence, the develop-
ment of methodologies for internal quality control was identified
as a major research issue by the scientific community (see the
editors' note in (Baltsavias et al., 2001)).
In the next section, we briefly review work on automatic road
extraction with emphasis on approaches dealing with urban en-
vironments and approaches employing internal quality measures.
In Sect. 3, we continue with a short overview of our extraction
system before describing details of the incorporated extraction
and evaluation methods in Sect. 4. Finally, the achieved results
are analyzed and discussed (Sect. 5).
2 RELATED WORK
Besides many user- or map-guided approaches, also numerous
automatic approaches have been developed (see articles in (Gruen
et al., 1995, Gruen et al., 1997, Baltsavias et al., 2001)). Most of
these efforts are directed towards the extraction of roads in rural
areas. Approaches designed to process satellite or low resolu-
tion aerial images generally describe roads as curvilinear struc-
tures (Heller et al., 1998, Wang and Trinder, 2000, Wiedemann
and Ebner, 2000) while those using large scale imagery (i.e., a
ground resolution less than 1 m) model roads mostly as rela-
tively homogeneous areas satisfying certain shape and size con-
straints (Ruskoné, 1996, Zhang and Baltsavias, 1999, Baumgart-
ner et al., 1999, Laptev et al., 2000).
Compared to the relatively high number of research groups fo-
cussing their activities on rural areas, only few groups work on
the automatic extraction of roads in urban environments. Here,
the road network is often modelled as a combination of grids with
a rather regular mesh size, i.e., the size of one building block.
(Faber and Fórstner, 2000), for instance, rely on directional in-
formation of lines extracted from scanned maps or low resolu-
A - 163
