the length as setting parameter (the direction of Y). Then, the
relative coordinate system (X, Y, Z) of the lattice point is
related to the absolute coordinate system (X, Y, Z) by the
Helmert transformation. A voxel-image is generated using a
projection that is the magnification conversion of an
independent rectification image according to an altitude value
obtained by resampling the lattice points of an absolute
coordinate value in the object space.
3.2.2 Extraction of rooftop and ground by horizontal
plane matching: The roof and ground are extracted using
multi-image matching with the voxel-image of the object space
using vertical direction searching, and the wall surface is
extracted with the normal direction by searching horizontally.
As shown in Figure 3, a voxel-image is divided into two
segments on a line segment of the roof footprint with the depth
direction of the wall. The rooftop and ground are searched in
the vertical direction to each division domain by using
horizontal plane matching with cross correlation to separately
extract the height. In the case of ground searching, the
occlusion obstructed more than half of the building in the multi-
view images. To address this, several images are sorted out for
ground searching to determine the front and back of the wall
based on the rotation direction of the 2D polygon’s vertex order
on each independent rectification image. The 2D polygon’s
vertex order is created by the projection transformation of the
3D polygon, which consists of the four corner points of the
initial wall surface.
3.2.3 Division matching for wall direction compensation:
The initial wall that is obtained from the building’s footprint on
an existing 2D map does not provide a good match to the actual
direction of the wall. Our proposed method compensates by
determining the precise position of the wall. The conceptual
diagram is shown in Figure 4.
BiRrsasanasdeses nus smencenee) CREAR
: O Matching Position
: «€ x» Searching Direction
2522 Compensation Line
MY
ee
Pe " E "
A 11 1 | X
Rtas
Matching in Voxel-Image
"assssssssssss
Regenerate of Voxel-Image
Figure 4. Compensation of Wall Direction
The proposed method can compensate for the wall direction by
using the results of cross-correlation matching for each section.
Cross-correlation matching is performed to detect matching
positions in the depth direction of the wall in each section of the
voxel-image, which is divided into lengthwise sections of the
wall. The position and direction of the wall are estimated using
the least-squares method from matching points of the
correlation peak position, and a new voxel-image is generated.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
This process is repeated several times, and when the amount of
compensation is below a set threshold, the wall surface
searching calculation is complete. In this way, an exact wall
position can be extracted.
4. EXPERIMENTAL PROCEDURE
4.1 Data Specification
4.1.1
Sensor Specifications and Flight Parameters: The
sensor specifications for the large format digital camera used to
take aerial photographs are shown in Table 1, and the flight
parameters to obtain multi-view images are shown in Table 2.
Number of Pixels
13,824 x 7,680 [Pixel
]
CCD Size 165.9 x 92.2 [mm]
Pixel Size 12.0 [um]
Focal Length 120 [mm]
Table 1. Sensor Specifications
Absolute altitude
Approximate 600 [m]
Over-lap Ratio
Approximate 66.7 [96]
Side-lap Ratio
Approximate 66.7 [%]
Number of Projects
2 [Projects]
4.1.2
Table 2. Flight Parameters
Experimental Data:
An example of the multi-view
images used for the experiment is shown in Figure 5.
xU
Figure 5. Multi-View Images using Experiments
The multi-view images are arranged as a complex image with
the partial clipping of images (11 images: from numbers | to
11) of the same building. Our proposed technique can
efficiently process multi-view images that are taken as a re
of regular and irregular flight paths.
sult
42 Exper
421 Ge
which are
segment of
voxel-imag
images. Mc
wall surfac
shows that
correctly £
polygon d:
processing
422 Ex
plane matt
the differe:
search for t
the pattern
is illustrate
processing
423 Di
Here wall
and it ver
direction «
position ot
result of th
evaluated q
5.1 Gener
We perfor
the roof fc
south. The
multi-view
line is take:
footprint li
figure, tow
of the part
figure, is c
rectificatio:
magnificati
vanishing
image is ci
to an altitu
and it accu
image. Ac
voxel-imag
dimensions
(one dimen
Next, the i
image is s
shows the
voxel-imag
[Image nun
building W
wall. In the
Figure 7, tl
the buildin
extract the
using the c
can be calc
