um 
the 
da 
dis 
Iter 
s in 
ir is 
line 
) is 
it is 
t to 
are 
the 
tion 
age 
ally 
ver, 
ons 
nap 
] to 
ant. 
it in 
two 
st to 
Ver, 
the 
1igh 
SEN 
und 
4 is 
are 
pad. 
d to 
line 
to 
ocal 
A new map is produced by the interpretation process. Each 
map point gets a label telling whether it was recognized in the 
image data or not. The map is inspected by the user. Larger 
continuious parts of the road network which not were 
detected may represent removed roads and should be further 
visually checked. 
3.2 Water Bodies 
The main pupose of detecting water bodies is to locate sea 
shore line segments that can be confused with road segments. 
The system includes interpretation of seas, lakes and rivers. 
For seas and lakes both water areas and the edges (sea 
shores) can be interpretated. Rivers are interpretated as darker 
lines against a brighter background. The interpretation 
process follows the same overall strategy as for roads. 
Segmentation of sea shores is performed by a filtering 
algorithm analog to the "template matching" described in the 
previous section, except that edges are extracted instead of 
lines. The water areas are segmented by statistical 
classification performed outside the AUTOSAT system. 
Rivers are segmented by line template matching on a negative 
image. 
3.3 Urban Areas 
À method, named line interference filtering, has been 
developed to recognize urban areas. The idea behind the 
algorithm is the assumption that recognition of the road 
network is crucial for visual recognition of urban areas. 
Accordingly, the method first extracts the road network from 
the imagery, then a filtering operation is performed extracting 
areas where the line density is high. The resulting image is 
finally postprocessed. 
The "template matching" method described in section 3.1 is 
also used for detection of roads in urban areas. After template 
matching noise reduction is performed. The line segments is 
then thinned to single pixel width. The thinning operator is 
implemented as an iterative filtering operator, using 32 filters 
of size 3x3 representing different patterns of line edges. 
During each iteration, the filters remove pixels first from the 
left, then from the right until the line thikness is one pixel. 
The algorithm preserves the center pixel of the thinned lines. 
Thinning of our result images is obtained within 3-4 
iterations for the test data. 
The idea behind interference filtering is to make an operator 
that generates some kind of interaction between lines with a 
response that is proportional to local line density. At least two 
lines must be present to create an interaction. This means that 
single lines, as single roads and the coastline, will not create 
any response. 
With a two-dimensional filter it is difficult to ensure that a 
line will not make interference with itself. To avoid this 
problem we designed one-dimensional filters, and these 
filters are run on only the parts of the lines that have an 
orientation that is close to orthogonal to the orientation of the 
filters. 
Figure 2. The four one-dimensional filters for extraction of 
(respectively, from left to right)) nearly horizontal, vertical, 
right diagonal, and left diagonal line structures. 
551 
The interference process is divided into three steps: 
1.  Decompose the line structures in the binary image into 
four new images containing respectively nearly vertical, 
horizontal, left-diagonal and right-diagonal lines. 
2. Generate interference between the nearly vertical lines 
with a horizontal filter, between the near horizontal lines with 
a vertical filter, and so on. 
3. | Combine the four images into one image. 
Step 1 is performed by filtering the binary image with the 
four simple filters in figure 2. Each filter extracts lines with 
orientation + 1/4 of the direction the filter is designed for. 
Hence, all directions should be covered by two orthogonal 
filters (e.g. the horizontal and vertical filter). However, we 
have included the two diagonal orientations to make a 
smoother combined image in step 3. 
Figure 3 shows the four simple interference filters. The filter 
response is equal to the number of lines covered by the 
filtermask. The interference filter operation is designed in 
such a way that a filter will not response if only the left or the 
right hand side of the center point of the filter covers lines. At 
least one line must be covered by the other part of the filter. 
The design ensures that the borders of the urban areas are 
preserved (defined by the outermost road). 
4. Detection of New Objects 
4.1 Roads 
A feature image generated by template matching as described 
in section 3.1 is used. The feature image is manually 
thresholded with a relatively high threshold value to limit the 
amount of noise. The recognition result is, however, not very 
sensitive to the value selected. The detection algorithm is 
using both a feature image and a binary thresholded image. 
In addition to roads the binary image will contain a lot of line 
segments that are not roads. Especially the coastline will 
appear very roadlike. The resulting map from map-guided 
detection of the water bodies is therefore used to remove all 
line segments having about the same position as the sea 
shores. In addition, lines inside larger urban areas, both old 
and the new areas detected as described in section 4.2 below, 
are removed. The road network inside urban areas is too 
complicated and detailed to be well recognized in imagery 
with a resolution of 20 and 10 m. The AUTOSAT system 
handles these areas just as urban areas without trying to 
extract any more information about them. However, small 
areas are not removed to limit the number of gaps generated 
in the main roads by removed urban areas. 
The resulting binary image still contains some noise. The 
noise will often appear as short lines in a more or less 
random pattern. Visually, one will often see that these lines 
probably are not part of the road network (the random 
orientation), and for the same reasons these lines will be 
removed by the system at a later stage in the recognition 
process. 
Figure 3. The four one-dimensional filters to generate 
interference between (respectively, from left to right), nearly 
horizontal, vertical, right diagonal and left diagonal line 
structures.