International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
changes of road networks in new/old maps with different scale
is shown as Fig.4.
(a) The old 1:50000 road map
(b) The new 1:10000 road map
(c) The change detection results
Fig.4 An example for change detection results using the double-buffer detection algorithm (In (c), blue thick lines represent appear
road lines, black thick lines represent appear new lines, the back-buffer and front-buffer detection distance is 0.67mm, the
change belief value is 0.9)
4. CHANGE DETECTION BETWEEM NEW IMAGE
AND OLD MAP
The main difficulty of automatic change detection for road
network between new image and old map lies in two aspects:
one is the tracing of unchanged road and another is the
extraction of new road. For detection and tracing of unchanged
old road, automatic detecting algorithms based on GIS
information are a good idea For the extraction of new road,
some new ideas and strategies including hybrid feature
grouping techniques, automatic road recognition based on
knowledge base, knowledge inference for road recognition,
road re-grouping etc. are discussed in the following.
4.1 Automatic change detection based on GIS
Extracting road networks from imagery with the help of
existing old maps can reduce the difficulty significantly. On one
hand, we can suppose roads change very slowly and this means
we can extract the road from images based on existent road
information. On the other hand, road network own perfect
topological properties and this means we can find another road
through connected road networks casily. Obviously existing
road in the old map can be employed as the start point of
finding and extracting road in new image. However, how to
detect the new road from images using old GIS information is
worth to discuss.
In images with low resolution, i.e., more than 2m pre pixel,
roads mainly appear as lines establishing a more or less dense
network, road tracing is not very difficult. Here we developed a
buffer detecting and tracing algorithm. The main idea of this
algorithm is to make a buffer for road in old map with a given
buffer distance. And in this buffer, query all extracted parallel
lines and link them according to the similarity of direction,
length of two lines. In this procedure, using buffer detection
algorithm for detecting road, four possible results can be
detected: road is not changed, road 1s changed and widen, road
is changed partly, road is disappear. The road can be detected
and extracted in this procedure under the guidance of GIS
information.
In high-resolution images, the appearance of roads is the mixed
appearance of all sub-structures including cars, zebra crossing,
lane lines and so on. Generally these structures can be taken as
the existence proof of road network and they can be used for
road extraction. But this is not enough. Because of too much
noise and influence in high-resolution images especially this
kind of proof is not exclusive, it is necessary to give some prior
462
information using old GIS data. A good example is the
extraction of bridges. The main strategy is to firstly detect rives
using GIS information because river change very slowly and
then search those candidate road segments vertical with river
central lines. At last these road segments are linked according
to some rules and the bridge will be extracted. Another
effective proof is cars in roads. Because of good properties of
cars, cars structures can be extracted easily and they can be
used for assistant in road extraction. It is difficult to directly
extract cars from images, but according to prior car models
from extracted roads, the difficulty will be reduced significantly.
An road extraction model based on cars was developed
(Sui,2002,2003).
4,2 Hybrid feature grouping methods for road networks
extraction
In middle-level vision, feature grouping is one of the key
techniques. Road grouping is the most important technique for
organizing short and incomplete road segments into long and
more reliable road segments. Hybrid feature grouping methods
includes two aspects: the grouping technique based on whole
relationship for low-resolution images and the grouping
technique based on profile tracing for high-resolution images
are employed.
The grouping technique based on whole relationship is that
many kinds of grouping contents including geometric properties,
image attributes and other direct or indirect information can be
quantitatively described [Sui, 2003]. Road grouping based on
this grouping strategy can be decomposed of two parts: one is
grouping similar road segments and another is the extended
grouping for non-similar road segments. Grouping similar road
segments means to group road segments in given range
(generally it is smaller value) according to similarity
measurement parameters. The latter mainly groups those non-
similar road segments that miss relative road information. After
these grouping procedures, the whole framework of road
network comes into being.
Opposed to the low-resolution images, in images with high
resolution, i.e., less than 0.50m, roads are depicted as elongated
homogeneous areas with more or less parallel borders. Because
the complex relation between roads and other objects, like cars.
buildings or trees, a reliable detection and extraction is often
difficult. Sub-structures of roads like zebra crossing, lane line,
white line, block line and other road markings have a strong
influence on the characteristics of roads. The appearance of
these sub-structures in images with high resolution are the multi
Interna
fragme
algoritl
princip.
For the
lines a:
vertexe
these p
until al
step, th
how to
For ew
length :
can sup
close tc
can be
length
informe
43 A
Structu
images.
informe
contras!
the alg
Under
supplen
statistic
viewpo
based o
best ch«
General
are mar
based o
practice
segmen
we ove
images,
from t
histogr:
these tl
Candid:
properti
e
Accord
the aut]
adaptiv:
Conside
model f
employ
features
sensitiv
kinds
features
and cur
