hr
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
uonooarp Sup[oo]
(a)
Figure 3: Appearance of highways in SAR images (a) Orientation dependent effects (b) Highways in reduced resolution, about 6 m
complement are layover areas smearing over the road. (E) Ad-
ditionally, context objects like bridges (also traffic signs, tunnels,
and vehicles) can be present.
In the same image but with reduced resolution, the fundamental
structure of a highway is emphasized (Figure 3(b)). It appears as a
dark, smooth-curved line, and the crash barriers are no more vis-
ible. However, a total corner reflection ("A" in Figure 3(a)) may
avoid the annihilation of the crash barrier at lower resolutions.
Hence, we restrict this part of the model to highway orientations
that deviate from the azimuth direction significantly.
3 ROAD EXTRACTION
3.1 Road extraction in rural areas
The approach used for automatic road extraction in rural areas has
been originally designed for optical imagery with a ground pixel
size of about 2 m (see (Wiedemann and Hinz, 1999), (Wiedemann
and Ebner, 2000)) and was adapted to SAR data (Wessel et al.,
2002).
The first step of the road extraction consists of line extraction us-
ing Steger’s differential geometry approach (Steger, 1998). It can
be performed in multiple images and with different parameters
settings for the individual road classes (highways, main roads,
secondary roads). In the next step, the lines of each extraction
are evaluated according to their road characteristics: curvature,
width, reflectance properties etc. Then, with the confidence mea-
sures thus gained, overlapping lines are fused using a best first”
strategy and a weighted graph of road segments is constructed
from the resulting (unique) set of lines. For allowing the elimina-
tion of gaps in the line extraction, candidates for supplementary
road segments are added to the graph - typically resulting in an
over segmented intermediate extraction. To extract the road net-
work from the graph, seed points are defined (i.e. high rated road
segments) and connected by optimal paths through the graph.
The union of these paths corresponds to the final road network.
3.1.1 Extraction with context objects The road extraction
algorithm allows to introduce additional segments together with
confidence measures (weights) on the basis of the above men-
tioned fusion. This property is used to introduce context objects.
At the time being, we assume that it would b possible to extract
these objects quite reliably. In the current state of the implemen-
tation, the extraction of local context objects is done manually
because, at the moment, the main task is to find out whether con-
text information is useful for road extraction or not. Advanced
studies for an automatic extraction of some of the context objects
can be found in the literature (Kirscht, 1998), (Kirscht and Rinke,
1998), (Barsi et al., 2002). For introducing context objects into
the extraction process it is important to (1) estimate the evidence
each context object provides for roads and (2) choose an appro-
priate representation form for each context object: High evidence
for roads is provided by context objects that almost exclusively
appear in conjunction with roads and rarely elsewhere. There-
fore, vehicles blurred in azimuth direction, and also bridges, get
high weights. Their representation form is a line. Other ob-
jects provide less evidence for roads. For example, alleys appear
nearby roads but also elsewhere. They are henceforth represented
as lines attached with low weights. Large traffic signs only appear
together with roads. However, their correct (automatic) interpre-
tation is assumed to be quite hard, so that they are added to the
graph as middle-weighted short straight lines. The same is true
for junctions, i.e. intersection points of roads. They are modeled
as low weighted points in the graph with several terminals that
allow connections between three or more lines.
3.1.2 Extraction with context regions Until now, context re-
gions are simply used to exclude urban and forest areas from the
extraction (Sect. 3.1). This was done because the computation
time increases with the number of potential road segments, which
is extremely high in the above mentioned regions. For this task a
urban-forest-mask was generated, which can be extracted directly
from the SAR data. A classification of X- and full-polarimetric
L-band data allows to extract rural, urban and forest areas, based
on the intensity values, ratios, and neighborhoods.
Furthermore, urban areas are now used as seed information in the
road extraction procedure. We introduce the contours of urban
areas as additional weighted segments in the same way as de-
scribed in Sect. 3.1.1. The evidence for urban areas to be a seed
point is very high. The advantage of introducing the contour line
is threefold. First, no further hypotheses or extraction attempts
of roads inside the city outline have to be made. Second, the
function of cities to link road parts together to a network without
interruptions by urban areas is fulfilled. Third, the contours are
especially helpful in the vicinity of urban areas because often, the
roads are not clearly visible in those regions.
3.1.3 Extraction of highways The extraction strategy for high-
ways consists of four different steps: (1) hypotheses formation in
low resolution, (2) hypotheses formation in high resolution, (3)
fusion of both resolutions, and (4) network generation. (1) To cre-
ate highway hypotheses in low resolution, dark and wide lines are
extracted (Steger, 1998). The resulting lines are weighted with
respect to highway construction parameters (width, length, cur-
vature). Especially variants of the Hough Transform for straight,
circular, and elliptical structures help to weight lines according
to their evidence being part of a highway. (2) In the high resolu-
tion, dark lines and thin bright lines are extracted, i.e., candidates
for the individual lanes and the crash barrier in between. To get
initial highway hypotheses, parallel dark lines enclosing a bright
line are selected. These line aggregation is rated according to
highway construction constraints and, in addition, according to
the gray value difference of the parallel dark lines. (3) All hy-
potheses are fused now using a "best-first" strategy. Thereby,
hypotheses extracted in both resolutions get the highest weights.
(4) Finally, the network is extracted by the graph-based grouping
algorithm described in Sect. 3.1.
Internat
The po
two tes
(AeS-1
Germar
tive eva
in (Wie
ually p!
ble 2, |
ness, ai
ated wi
uation |
for mai
correct
Figure
without
Introdu
tions (I