570
,
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
The walls were scanned on a cloudy day to ensure as much
diffused illumination as possible. However as shown in Figure 4
this was not entirely successful. Walls 1 and 2 contain shadows
thrown by a nearby tree and a nearby pillar respectively.
5.1 Walll
The mortar channel in Wall 1 is about 10 mm deep. Each brick
is 220 mm wide and 70 mm high. At an average point spacing
of 3 mm this yields about 2000 points per brick. Therefore,
invalid bricks, i.e., incorrectly segmented bricks should have a
point count considerably less than this. A value of 750 (slightly
less than half a brick) was used in the tests. The rgb triplets of
each point was first converted to hsv, see Figure 5. The
following point attributes were used, Hue, Saturation, Value,
intensity and mean distance to neighbouring points.
Figure 5 Point hue (left), saturation (middle) and values (right)
of wall 1. The value component suffers the most from the
shadow (dark area) cast by the tree.
The edge strength was computed as the product of the point
attribute at both edges or the length of the edge. Attribute
difference was also tried, but the attribute product provided the
best discrimination between brick and wall points. Once the
edge strengths had been set the edges with a value below a
given threshold were removed. The best value for the threshold
was obtained from the mode seeking technique described in
section 4.7.
As shown in Figure 6 the segmentation is most successful when
edge length is used as the edge strength criteria. This is helped
by the fact that the mortar channel is both deep and wide and
the scan resolution is fairly high. The results of the hsv
segmentation are fairly similar with the value segmentation
possibly performing worse than the hue and saturation
segmentations. The hsv segmentations yield at least 5 over
segmentations on the left side of the wall. This is because the
segmentations were optimized for the right side of the wall.
In Figure 7 is shown the means by which the optimum threshold
is selected. In this example invalid segments have a point count
of 750 or less, and the edge strength is based on edge length.
The images show the segmentations at different points on the
graph. Note that as expected the optimum result occurs near the
peak of the curve.
5.2 Wall 2
Wall 2 presents a more difficult problem and is typical of the
brick detection problem trying to be solved. The mortar channel
is both shall and narrow. Moreover the granite bricks are of
different shapes and sizes. Here invalid bricks were chosen as
containing a 750 points or less. The rgb triplets of each point
was first converted to hsv, see Figure 8.
mmm
ë
wmm mm m
SSnan *
aft m
«»I wmm .' »«
,*am»
* imi *
m ,
Using edge length
I Vi *
lU w 4 .
Using value
: Cft «
Using saturation
iilSfc
E.1.U - ftr/1
Using hue
Figure 6 Segmentation using different edge strength criteria.
* ■ *■
mm u. 1 /
mmMH . ■.
(ME* «#% •
>
m # **' .**<•
•r.a.fcEM' ■
«MW«» :
mrv',,.. ..i.
fl t WKtM .
"*№ '
•spiff»«
m mmWMgm
jCzti 1 » J
pw *
' wA .
.
Under
Optimum at
Over
segmentation at
threshold 5.8 mm
segementation at
threshold 5.2 mm
(at peak of curve)
6.5 mm (right of
(left of peak)
peak).
Number of components vs. Edge strength threshold
140
Edge strength threshold
Figure 7 Selection of threshold for segmentation.