You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Stilla, Uwe

In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
Anne Grote, Franz Rottensteiner
Institute of Photogrammetry and Geoinformation, Leibniz Universität Hannover, 30167 Hannover, Germany
(grote, rottensteiner)@ipi.uni-hannover.de
Commission III, WG III/4
KEY WORDS: High resolution, Aerial, Urban, Automation, Extraction
In this paper, a road extraction approach for suburban areas from high resolution C1R images is presented. The approach is region-
based: the image is first segmented using the normalized cuts algorithm, then the initial segments are grouped to form larger
segments, and road parts are extracted from these segments. Roads in the image are often covered by several extracted road parts
with gaps between them. In order to combine these road parts, neighbouring road parts are connected to a road subgraph if there is
evidence that they belong to the same road, such as similar direction and smooth continuation. This process allows several branches
in the subgraph which is why another step follows to evaluate the subgraphs and divide them at gaps which show weak connections
after gap weights are determined. A digital surface model, if available, is used in the grouping and road extraction step in order to
prevent high regions from being extracted as roads. The results of the road extraction with and without the digital surface model are
compared in order to show how the extraction is improved by the surface model. It also shows what can still be expected from the
extraction if no digital surface model is available.
Roads are a very important part of the infrastructure, especially
in urban areas. Road data are used in many applications, for
example car navigation systems. For these applications it is
important that the road data are up-to-date and correct. As the
road network is subject to change, especially in suburban areas,
the road databases have to be updated frequently. This is often
done manually with the help of aerial or satellite images. In
order to reduce the costs and the time required for map
updating, it is desirable to use automatic procedures for the
extraction of roads from these images. Today, roads are to a
large degree still extracted manually, especially in urban areas,
because of the relatively high complexity of urban
environments compared to open landscapes. For open
landscapes, road extraction algorithms that are reasonably
reliable already exist, e.g. (Zhang, 2004). This was confirmed
by the EuroSDR test on road extraction (Mayer et al., 2006). In
this test, several state-of-the-art methods for road extraction
were compared, using imagery with a resolution of 0.5-1.0 m.
The results were reasonably good in rural scenes of medium
complexity, but the algorithms did not perform well in urban or
suburban areas.
There are many different approaches for road extraction from
optical imagery, and in recent years the number of those that
deal with urban areas has increased. Road extraction algorithms
can be classified into line-based approaches and region-based
approaches. Line-based approaches, which model roads as one
dimensional linear objects, are mainly used in open landscapes
with images of middle to low resolution, and they are not
suitable for urban areas. An approach for urban areas that
extracts middle lines and edges of roads and groups them to
form road lanes using aerial images of very high resolution (0.1
m) is described by Hinz (2004). In most other approaches
regions are extracted from images with a resolution of
approximately 1 m. One example is (Zhang and Couloigner,
2006), where a colour image is classified and the regions
classified as roads are refined in order to separate roads from
false positives such as parking lots. Another example for a
region-based approach is (Hu et al., 2007), where footprints of
roads are extracted based on shape, and the roads between the
footprints are tracked. The high complexity of urban and
suburban areas makes road extraction from greyscale aerial
images without further information difficult because many
different structures in urban areas have an appearance similar to
that of roads. Therefore, most approaches use additional
information, for example colour (Zhang and Couloigner, 2006;
Doucette et al. 2004), Digital Surface Models (DSMs) (Hinz,
2004) or both (Hu et al., 2004). Information about the position
of roads from an existing road database can also be used, e.g.
(Mena and Malpica, 2005). Prior information about the road
network is another possible source of information. Price (1999)
assumes that the road network forms a regular grid. This is also
done by Youn and Bethel (2004), though they use less strict
requirements for the grids.
In this paper, a region-based approach for road extraction from
aerial colour images with a resolution of 0.1 m is presented.
Optionally, a DSM can be used as an additional source of
information. Apart from the DSM, our approach does not
require other sources of information such as an existing
database, as used in (Mena and Malpica, 2005). Since we work
in suburban areas, the approach does not rely on particular
properties of roads like road markings, as used in (Hinz, 2004)
or a regular road grid, as used in (Price, 1999), and all roads
should be extracted, not only major roads. In the approach, an
image is first segmented and then road parts are extracted from
the segments. These road parts are assembled into road
subgraphs. In this way, there is no need to assume that a whole
road can be extracted undisturbed. The subgraphs can contain
different branches which represent different hypotheses for the