In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
Figure 6. Left: Airplanes which were detected as buildings in
(1D2), Right: Elimination of airplanes with (1H2) D4.
((Ifl2) D4) U 3: Shadow regions on buildings are replaced with
building regions and by this combination (Figure 7). Since
method 3 brings the buildings which could not be detected well
by method 4, and method 3 is not influenced by shadow, this
combination provides better completeness (in Figure 12, R3).
Figure 7. Left: buildings without the regions which covered by
shadow in ((1H2) H4), Right: more complete roofs with ((102)
04) U 3.
After the union process with the results of the Method-3, the
vegetation on the roof tops is still a problem. Intersection of the
nDSM and the NDVI algorithms provides the tree and
vegetation regions on the roof tops. Intersection of the extracted
vegetated regions with building polygons of the Method-4
results in the roof regions which contain vegetation (Figure 8).
Figure 8. Roof regions which contain vegetation.
After adding the roof regions which contain vegetation into the
detection result (in Figure 12, R4), the correctness and
completeness values are 85% and 7%. As mentioned before,
since method 2 have detected all buildings although not fully,
the final building polygons should overlap the results from
method 2. If the building polygons of result (R4) do not overlap
with the results of method 2, they are eliminated. The
correctness of the results is improved to 91% and the omission
is 7% (Figure 12, R5).
5.2. Using edge information for improvement of correctness
Image data provide edge information, and this can be used to
find the precise outlines of the buildings. Firstly, the Canny
edge detector (Canny, 1986) has been applied on the
orthoimages. The edges have been split into straight lines using
comer points which were detected by comer detection (Harris
and Stephens, 1988). This has been performed using the
Gandalf image processing library (Gandalf, 2009). The straight
lines which are smaller than 1 m. have been considered as noise
and they have been deleted. The straight lines which may
belong to building outlines have been selected using the outline
of the detection result (which comes from the combination of
methods) and a 2m buffer zone (lm inside, 1 m outside of the
building outline). If the straight lines are neighbours in the
buffer zone, the longest straight line has been selected. There is
an exception for this neighboring criterion: the start or end point
of a straight line should not be the closest point to the
neighbouring line. With this exception, we avoid the
elimination of lines, which are almost collinear (Figure 9).
Figure 9. Left: the straight lines which may belong to the
building outline (yellow) and Right: long lines (red).
After selection of the straight lines, they have been converted to
closed polygons. For the conversion to polygons, a sorting of
the lines in clock-wise direction is used. To perform sorting, the
travelling salesman convex hull algorithm (Deineko et al.,
1992) has been applied. After closing the polygons, we
separate the lines into those that were detected from the images
(red) and the ones added by this algorithm (blue) (see Figure
10). The red straight lines, which are shorter than 10 m. and
form an acute angle (between 1 and 80 degrees), are eliminated
(Figure 10), as well as all blue lines.
Figure 10. Line elimination procedure when the line length is
shorter than 10 meters and has acute angle with its
neighbouring lines (red: eliminated lines, blue: lines added by
the travelling salesman algorithm, yellow: acute angle).
If two red lines form an acute angle and are shorter than 10 m.,
then both lines are eliminated. After this elimination, the
travelling salesman convex hull algorithm has been applied
again using the non-eliminated red lines and generated the
refined building polygons (Figure 11).
Figure 11. Final building polygons (yellow)., and reference data
(red).