mbul 2004
results
nt appear
67mm, the
le is the
etect rives
lowly and
with river
according
Another
operties of
ey can be
to directly
ar models
enificantly.
developed
networks
f te key
hnique for
) long and
g methods
on whole
grouping
on images
Np is that
properties,
ion can be
based on
rts: one is
+ extended
milar road
ven range
similarity
those non-
tion. After
« of road
with high
elongated
s. Because
„like cars,
n is often
, lane line,
€ 3 strong
earance of
e the multi
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
fragmented, broken parallel lines. Aiming at this, a new
algorithm called profile-tracing algorithm is presented. The
principle of the new algorithm consists of two steps:
eo Firstly group these lines for creating possible
candidate road segments according to profiles;
e Detect changes of road and extract road sides
according to the candidate road segments.
For the first step, the main strategy is to select reliable parallel
lines as searching start lines, and make profiles at their two
vertexes so as to find possible conjoint parallel lines. Then link
these parallel lines into one long line and repeat this procedure
until all the given conditions can't be satisfied. For the second
step, the key problem is whether roadsides can be extracted and
how to trace roadsides according to the existence information.
For every road segment, maybe it doesn't overlay the whole
length of actual road. However, for many road segments, they
can supplement each other and form a new line whose length is
close to the whole road length. Base on these lines, roadsides
can be created with the help of other grouping lines. The road
length and width can be defined with the help of GIS
information.
4.3 Astrategy for extracting road networks from images
Structure features are main representation information of
images. Almost all road extraction algorithms employ structure
information. However, when there exist strong noise, low
contrast and other factors in images, the extraction results using
the algorithms based on structure features are often not well.
Under this situation using statistical information is a good
supplement. But how to define a statistical model and utilize the
statistical parameters are key problems. Obviously from the
viewpoint of extraction strategy, the hybrid extraction strategy
based on structure information and statistical information is the
best choice.
Generally, a lot of statistical models are very complex and there
are many unknown parameters to be defined. So the algorithms
based on these statistical models are not practical. A simple and
practical statistical model is necessary. In general, local road
segments in images have better statistical properties. Suppose
we overlay a window with a given window size on the road in
images, the local statistical properties of roads can be reflected
from the properties of the gray histogram, gray gradient
histogram, gray gradient intensity histogram in the window. So
these three parameters can be used as statistical parameters.
Candidate road segments can be found if the statistical
properties in this window satisfy the following three conditions:
e The grey histogram in the window has only one peak
eo The gradient direction histogram in the window has
two symmetric peaks
eo The gradient intensity histogram in the window has
one no-zero peak
According to the above statistical properties of road segments,
the authors design a new road-tracing algorithm based on the
adaptive template (Su1,2002).
Considering the limitation of traditional Snake [Kass et al.,1988]
model for processing the road with lager road width, the authors
employ the Ribbon-Snake model for extracting the road
features. To avoid slow convergence properties and high
sensitivity to initial position of traditional Snake Model, five
kinds of condition. including continuity, curvature, image
features, shape, range are involved in this model. The continuity
and curvature conditions are the same definition with the Snake
463
model. The image features consider not only the gradient
parameter but also the texture information. The least squares
algorithm is employed for the shape constraint model. The
range condition mainly control the selection of initial position
in images. Here we use the buffer created by traced road central
lines and road width as the searching and restriction range.
4.4 Automatic road recognition based on knowledge base
Automatic road recognition is a necessary task for extracting
road networks. After many years of research, people realized
that it is impossible to design a common algorithm for
recognizing road networks from all kinds of images and it is
also not enough to only depend on image information for
automatic recognition and extraction. After great efforts of
many years, many researchers have waked up to successful
extracting of road networks should recur to the knowledge. In
this paper, the authors built a road recognition expert system.
All the knowledge is put into the knowledge base.
Generally three obstacles hold back automatic recognition: One
is how to acquire recognition knowledge, another is that how to
represent knowledge quantitatively and the third is how to use
knowledge for reasoning and recognizing road. In this
framework we first define a basic model for road. Main
parameters including grey value, grey variance, the length,
width, curvature are involved in the road model. The fuzzy
theory is employed for representing road models. According to
different parameters we define a series of different fuzzy
functions for road model. And a road recognition expert system
is built. All kinds of road recognition knowledge are all put in
the knowledge base. We use fuzzy production rules for
representing road knowledge and uncertain reasoning models
for road knowledge reasoning. All the recognition results are
represented as belief values. The road segments whose belief
value is greater than the given threshold are extracted as road.
4.5 Re-grouping road based on global information and
creation of new road
After automatic road recognition the whole framework of road
network is built and the hidden semantic contradiction is solved.
However, the work of road extraction is not finished. The road
network is still incomplete because of the following reasons: the
big road gaps caused by rows of trees and buildings, the wrong
link among road segments, the missing road caused by wrong
recognition and other factors. So it is necessary to further
extract and link road network. The principle and the strategy are
to use the extracted road segments and the global information of
road network for guiding road extraction. On one hand, the
existent road can be used as the start point for searching and
linking road segments; on the other hand, different road
segments can be linked according to some rules like distance,
grey value etc. Obviously road re-grouping is not blindfold
when high-level information feed back to middle-level or/and
low level processing.
Under the guidance of global topological information, the road
grouping and linking the iterative procedures until the last road
networks are created. The authors think four parts should be
processed in this procedure: road re-grouping, creation of road
crossings, creation of road branches, iterative road extraction.
Some extraction results based on the whole framework in this
=
2.
e
paper are shown as Fi
