UPDATE OF ROADS IN GIS FROM AERIAL IMAGERY:
VERIFICATION AND MULTI-RESOLUTION EXTRACTION
A. Baumgartner!, C. Steger?, C. Wiedemann‘, H. Mayer‘, W. Eckstein?, H. Ebner!
!' Lehrstuhl für Photogrammetrie und Fernerkundung
?Forschungsgruppe Bildverstehen (FG BV), Informatik IX
Technische Universitát München
Commission III, Working Group 2
KEY WORDS: Image Understanding, GIS-Updating, Road Extraction
ABSTRACT
Aerial imagery is an important source for the acquisition and update of GIS data. By using digital imagery it is possible to automate
some parts of these tasks. In this context this paper proposes a new approach for the automatic update, i.e., verification and extraction,
of roads from aerial imagery. The verification process evaluates road axes from GIS data based on the analysis of profiles taken
perpendicularly to the axes. It is possible to handle inaccurate axes, as well as to detect initial points for branching roads. The process
for the extraction of roads is independent of the GIS data, but relies on knowledge about roads provided by a road model. This
model comprises knowledge about geometrical, radiometrical, topological and contextual properties of roads at different resolutions.
Multi-resolution extraction is applied because distinct characteristics of roads can be detected best at different resolution levels. By
fusing results of different resolution levels the distinct characteristics of roads are integrated. Examples for the verification as well as
the road extraction are given.
1 INTRODUCTION AND OVERVIEW
Data capture and update are very important tasks to improve
or preserve the value of data in geographic information systems
(GIS). Update is equivalent to the verification of old data and
the extraction of new objects which have to be integrated into
the GIS. It is usually done manually by an operator and is time
consuming and expensive. Therefore, a lot of research work is
dedicated to the development of more efficient ways for update
of GIS data. Research on the automatic extraction of man-made
objects, like buildings or roads, from aerial or satellite imagery has
been carried out since the seventies, e.g., (Bajcsy and Tavakoli,
1976). In the beginning the attention was focused on automatic
data-capture for maps and GIS. However, the support that GIS
data can give for the interpretation in the context of update was
only realized recently.
Whereas a lot of work exists on the extraction of roads, which
is the type of object that is dealt with in the remainder of this
paper, relatively little work has addressed the verification. The
verification scheme described in (Plietker, 1994) is based on the
extraction of edges close to and parallel with the given road axes.
If a certain percentage of the edges, i.e., hypotheses for roadsides,
can be matched to the road axes, the GIS data is assumed to be
correct.
For the extraction of roads methods like profile matching and
detection of roadsides are used. The approaches vary in the way
how different methods are combined as well as how additional
knowledge, e.g., geometrical constraints, is incorporated. A main
criterion to distinguish the works is the interaction of a human
operator. In semi-automatic schemes an operator selects an initial
point and a direction for a road tracking algorithm (McKeown
Jr. and Denlinger, 1988, Heipke et al., 1994, Airault et al., 1994,
Vosselman and de Knecht, 1995). In (Gruen and Li, 1994) the
operator marks a few points of a road segment and a dynamic
programming based algorithm finds the road which connects these
points. This is advantageous because the path of the road is
more constrained and a more reliable handling of obstacles is
possible. A similar approach based on so-called “ziplock” snakes
53
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
is given in (Neuenschwander et al., 1995). A fully automatic
approach is presented in (Barzohar and Cooper, 1995). Stochastic
methods are used to find seeds for the road extraction. Roads are
found based on a grey level model and on assumptions about
the geometry of roads by dynamic programming. In (Ruskoné
et al., 1994) seed points for the extraction of the road network
are centers of elongated regions found by a segmentation. Based
on the elongated regions and their directions, road segments are
extracted using the homogeneity of the road surface. In order to
extract the road network geometrical constraints are taken into
account, and hypotheses about connections between single road
segments are checked.
This paper proposes a new approach for the automatic update
of roads from aerial imagery. The verification of roads employs
a simple model based on the analysis and tracking of profiles
taken perpendicularly to the given GIS axes. Strong edges in the
profiles are linked and checked for colinearity, parallelism, and
their distance to the GIS axes. The result distinguishes verified,
inaccurate, and rejected GIS axes as well as initial points for new,
branching roads. A detailed description and results are given in
section 2. In section 3 the automatic extraction of roads from
aerial imagery is described. It is independent of GIS data but
uses a more detailed road model incorporating different kinds of
knowledge about the characteristics of roads. Due to the fact that
different characteristics of roads can be detected best at differ-
ent resolution levels, evidence for roads is extracted at different
resolutions. The original image has a ground resolution of about
25 cm. To detect roads as homogeneous areas with parallel edges
it is slightly smoothed to reduce the effect of noise and small dis-
turbing features (high resolution). In an image reduced to a scale
where roads are only a few pixels wide (low resolution) road axes
are extracted. A combination step fuses both results. The result
of the fusion step is taken to direct the search for road markings to
get more evidence for the roads. Road markings are very weak in
the images and therefore the image is nearly not smoothed when
extracting them. This only gives reasonable results because the
place where to search for is known a priori. Finally in section 4
conclusions are given.