XXIX-B3, 2012
)057 JAPAN
Iti-view images
ssive earthquake and
h simulations require
. Progress in sensor
ies are inadequate. In
1g is complicated, we
1 the stereo-pair. The
mages is the focus of
al advantages, and its
complicated in its
ture searching with
n the 3D coordinate
pose an improvement
rocess, the extraction
cult because an aerial
ated edge features in
it. We propose an
atching, mismatching
> of an image pair is
| image caused by the
ents. We propose an
:D TECHNIQUE
tion of building walls
ent rectification. This
line segments of à
> top horizontal edge
led edge matching in
/e demonstrate the
hh in addition to the
ftop and the ground-
ssing flowchart of the
fication
indicates the method
image) generated by
ed IR method in this
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
Aerial Images
(Multi-view images)
Independent
Rectification
Edge Extraction umm
de TL
/ Edge Image f| Edge Direction y.
Independently
Rectified Images
Object-Space
Searching
Edge-image
Matching
HEE BEERS Wraps aa sae ERR
== Éreationof
- Voxel Image |
Voxel Image
(Image Matching)
|». Wal [| Horizontal
= Matching || Matching
sassseMessssssusessssslkueunsususs
TE
/ Wall // Roof // Ground /
Figure 1. Flowchart of Proposed Technique
N DT TE
"hanumunzuuzuruuz(jaumumnuua
Gm EEE EERE EERE EERE ERE BE HERE R
BHIRBAABID ERB RE msssssss
QoSusssaszsssssesasostusesttslasus
GGBOEOWCRO E OW WO GE OW ELUR NDOD RO ROGEORDORO GRO GEOUDOEOXD E GEO OG
Candidate
of Building
Footprint
X 1 0 0 cos, 0 sin ©,
Y |=/0 cos 0, -snQ, 0 1 0
Z 0 3snQO, cos Q, -sin®, O0 cos 0,
coK, -snK, 0 x X,
sinK, csR,.0 y dv Y, (1)
0 0 1 -f 2,
Next, we define the horizontal virtual-base plane as the rotation
angle of the camera position (= omega, phi, and kappa) as 0 and
leave the other camera parameters, such as the camera-lens
position and focal length of the projection central point,
unchanged. The four corner points of the exterior-oriented
image of the absolute object-coordinate system are projected
onto the horizontal virtual-base plane according to the
collinearity condition. In this way, we can obtain the four
corner points that are converted to the photograph-coordinate
system on the horizontal virtual-base plane. We can obtain IR
images by performing a projective transformation of all the
multi-view images onto the horizontal virtual-base plane using
this method. After this projection conversion, the IR images
satisfy the collinearity condition, and all pixels of the multi-
view images correspond to the original images.
3.2 Plane extraction by object-space searching
In the existing technique, occlusion frequently occurs between
multi-view images, and the feasibility of matching decreases on
planes with heavily distorted images. Object-space searching is
implemented to solve this problem. In the three processes
shown below, multi-image matching is performed for planes of
different directions. Rooftops, the ground, and walls are
matched separately in the object space. Finally, we can try to
extract the building shape based on plane matching.
3.2.1 Generation of Voxel Image: To perform plane matching
in various directions in the object space, a voxel image is
generated. The conceptual diagram is shown in Figure 3.
A Object-Space Coordinates
w........”” ^
: IR-image :
© Camera Point
OQ Feature Point
Figure 2. Independent Rectification Method
First, to generate concrete — multi-view images, the
corresponding points are computed using coordinate
transformation from the relative photographic-coordinate
System to the absolute object-coordinate system, as shown in
formula (1). Using formula (1), the four corner points of the
exterior-oriented image in the relative photographic-coordinate
System are transformed into four points in the absolute object-
coordinate system, accounting for the camera lens position and
the camera rotation angle. In this formula, & indicates an image
number (1, 2... N: N is the total number of multi-view images), f
indicates the focal length of the aerial digital camera.
Roof Outline
Helmert
transformation
(0, 0, 0)
Absolute
coordinate system
Figure 3. Generation of a Voxel-Image
The range of a voxel-image determines the size of the bounding
box including line segment candidates on the roof footprint,
which consists of the top edges of the wall. The width of the
bounding box is determined by the length of the line segments
of the roof footprint (the direction of X), the height is
determined by the length between the rooftop and the ground
(the direction of Z), and the depth is determined by searching
