XXXIX-B3, 2012
BUILDING PPO
ur information and so
nplex texture or colour
it is reasonable to use
e higher above their
hod for this goal.
th planar and vertical
ble image in the same
] from DSM only give
of the buildings. The
ted to be extracted from
VI, called masks, are the
ects into buildings.
+
ky
OR
Fs e
x PA
f»
%
n.
Pj.
v
y
wg
Pe
nterpretation
image is employed to
O), which should only
ks segmented by MCW
to a building footprint.
objects from the PPO
rated PPOs compose a
accuracy in terms of the
t. Generally it refers to
egion. For example, an
n type of tree, or area
dium resolution remote
xture of the spectrum of
igh resolution imagery
ognize the pure pixels.
get, most pixels can be
ain class, such as roof,
, tree, water, or band.
ot for building structure.
identical pure pixels as
(PPO), and process the
ie computation load and
do this, we employed a
d segmentation named
gnition merges a pair of
issimilarity in sense of
hen pixel is taken as the
vill be grouped as many
threshold gives larger
iber of them. Of course,
re likely “pure”. RGB
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
image with 20cm resolution generally require 10 to 20 scale
threshold to get relatively PPOs.
PPO grouping mask:
Building blocks segmented from DSM are taken as masks to
integrate the PPO to a building footprint. Mask usually has
fatter area than the building in RGB according to our
processing. Some other PPOs like street or ground may be
covered or partly covered by it. So the mask should be eroded
to a smaller area to make sure that each object covered by the
mask is a part of the building.
Tree eliminating:
There always are some small trees planted along the walls of a
house in Japan. The PPOs filtered by building mask sometimes
include one or two of them. A threshold of NDVI is used to
remove these objects. The remaining building PPOs are
grouped to compose a building footprint which has quite high
accuracy in the sense of both shape and location.
Region fullness processing:
In tree eliminating, some PPOs of building may be taken as tree
objects and be excluded. This error can be fixed by region
fullness processing. That is, a fullness measure is defined as the
ratio of the building footprint to the building mask. For those
under a certain value, object based region growing is
implemented. The neighbouring PPOs covered by the mask
within certain hue difference will be taken as building PPOs.
4. 2D MODELLING
Objective: a polygon for a building. The as-is polygon model is
subject to a square constraint, thus the modelling process
becomes a problem of optimization. Neighbouring lines of the
polygon are perpendicular to each other in the modelling
process. The line is fitted to the maxima of the intensity
gradients. Corners are generated initially by corner detection
and then dynamically added or merged during the iteration. The
whole procedure is entirely automated.
4.1 Main orientation pair
Square constraint.
The pair of main orientation of a building is represented by
angles auf. and subject to |a — A |= 90;a, B € (90,90). It
is observed that houses are usually built along a certain
orientation for a natural neighbourhood. Almost all the studies
consider this in their own ways. Primitive based methods
directly use models subject to the principle in the procedures.
Many others fit the line features with this constrains. We adopt
the latter idea since we have corners and lines extracted and
connected from the image.
District based orientation estimation:
A house’s orientation is estimated according to the whole
neighbourhood rather than to a singular house. In direction
fitting for a single house, the corners can be used to estimate are
relatively less. Sometimes the direction estimation has obvious
error. For a region with many houses, the direction errors are
not uniform, so that the models display a mass. Block direction
pair clustering (BDPC) is developed to classify the
neighbourhoods into several groups. Neighbourhood orientation
is estimated relatively and assigned to each house belonging to
It.
BDPC algorithm can be described in Fig. 3. The distance of a
pair of buildings is defined as the shortest distance between the
43
contours of the two buildings. The distance of each pair of
buildings is calculated. The district is grouped according to the
cluster which is computed on distance matrix. The main
orientation of a district is estimated using all the slopes of the
lines in this district by angle histogram. There will be two peaks
for a district, which has difference of 90 degree.
Buildings |
AES
| Contour distance matrix |
Clustering |
pr
Districts |
e eee
Orientation estimation for
each District
Y
Modeling for each
building
Figure 3 main orientation estimation
This technique increases the accuracy significantly, because
larger samples give more robust estimation. Fig 4 shows a
modelling result for a test sample district naming datal using
BDPC. Fig. 4 shows the building PPO (a) and the groups of the
district (b) in various colour. The district division is not so
correct for the houses arranged as one row in the middle of the
scene. Some of them are falsely classified as their neighbouring
blocks, so their orientation estimations are wrong. Fig. 5
displays the detected district orientation by the same colour
with that in Fig. 4 superimposed on the angle histogram
respectively. Each of the figures shows approximately two-
peaks distribution of the angles, which different by about 90
degrees. The two peaks corresponding to the main orientation
couple for this district.
Figure 4 District grouping
