tanbul 2004
| new road
ire 2). Then
to consider
shifting the
of matched
dow
dated
nd matched
ng and the
ast squares
on a road
is achieved
e 10).
ie complete
zments. To
1endly post
diting road
te matching
2 segments.
lection and
segment.
tching fail
hen deletes
serting and
sed on line
a user will
must click
» extracted.
is is indeed
alse alarms
part is the
user-given
ond part i$
ff building
- rectangles
> matching
ndaries are
; applied to
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
3.1. Line Extraction
To extract building boundaries, we first extract lines from the
image. A subimage of 100 x 100 pixels is defined centred on
the user input position. For line extraction, we used the
algorithm proposed by Burns et. al(1985).
The following images are an example of the subimage centred
on a building roof and lines extracted.
Figure 3. A subimage on a building and line image
3.2 Estimation of Building Position and Orientation
The orientation and position of the building within the
subimage are estimated by the lines extracted. Building
orientation is estimated first by voting the orientation of
individual line segments. The most popular angle is assumed as
the initial value of building orientation.
The estimation of building position is also done by voting. This
time, we first slice the subimage into small linear sections along
the direction of building orientation (see figure 4) and vote the
number of line elements within each image slice. The slice with
the maximum vote is assumed as the initial position where one
long side of a rectangular-shaped building is located.
— Building
ma =
: orientation
Voting
on image
slide
Figure 4. Estimation of the position of a building
Once we determine the initial position of building position and
orientation, we refine these values by least squares template
matching as the one describe in section 2. We first define a line
template whose orientation and position are set as the initial
values (see figure 5). Again, we assume the relationship
between the line template and the true line element from a.
building as similarity transformation. The equations 1 and 2
also hold in this case. Based on these transformations we can
design a least squares template matching. Through matching,
we refine the position and orientation of the line element and
hence those of a building to be extracted.
After the matching, we then have identified the position and
orientation of one long side of building boundaries. We name
this as the first long side. We can then estimate the position and
orientation of the other long side (second long side) of building
by slicing the line image with the new orientation and finding
the slice of the largest vote located opposite to the first long
side with respect to the subimage centre (see figure 6).
— | orientation refinement
Line Template : iy
position refinement
Figure 5. The line template and template matching
Subimage Centre
^.
yn long side
| —- Second long side
\
\
Figure 6. Estimation of second long side
Next, we need to find the other (short) sides of a rectangular-
shaped building. For this purpose, we again slice the line image.
But this time, we slice the image along the direction
perpendicular to the orientation of the long side. And we slice
not the whole image but only for the part between the first and
second long sides. The first short side is determined as the slice
with the largest vote. The second short side is determined as the
slice with the largest vote among those lie opposite to the first
short line (see figure 7).
In reality, however, it is not uncommon that short sides of a
building produce very weak edge patterns. As a result, lines for
those parts are often not detected at all. The line image we used
here also has this problem. We cannot see meaningful line
edges from the short sides of the building. Although we devised
an algorithm to extract lines from the all four sides of a building,
the location of short sides would not be very correct. Figure 7
shows this problem.
Nevertheless, we can retrieve the orientation and position of
long sides quite accurately. Instead of devising a more
sophisticated algorithm to correctly locate short sides (which is
“virtually” impossible), we have assigned this task to human
operators. With relatively simple operations such as rotation,
scaling and translation, we can edit the building rectangle to
correctly describe the underlying building. Figure 8 explains
this process.
3.3 Template matching of a building rectangle
In most urban areas, we can observe a collection of buildings
with a very similar shape. Typically in Korea, the apartment is
the most popular and preferred type of residence. There are
indeed many large apartment complex in any major cities in
Korea and all apartment buildings do have a very similar shape
with each other.
For such cases, building extraction may be done easily if we
can use the previously extracted building rectangle. For this
purpose, we use the least squares template matching as before.
We define the building rectangle extracted previously as the
template of the building to be extracted. Again, we can assume