nt
E EE
AUTOMATIC BUILDIND DETECTION FROM HIGH RESOLUTION IMAGES BASED
ON MULTIPLE FEATURES
Zhen Ge. Qiu, /EFE Member” *, Shi Tao. Zhang”, Chun Ling. Zhang’ , Jin Yun. Fang?
“ The Institute of Computing Technology of Chinese Academy Of Sciences, No.6 South Ke Xue Yuan Road, Beijing,
China - qiuzhenge@ict.ac.cn
" University of Science and Technology Beijing, Beijing, China -
changgong-ok@263.net
* Surveying &Mapping Bureau Of He Nan Province China
Commission II, IC WG II/IV
KEY WORDS: Edge detection; Earth Mover's Distance; Curvature; Fuzzy density; Hough transform; Building Detection;
ABSTRACT:
Automatic recognition and reconstruction of buildings from aerial and space images is of gre
applications such as cartography and photo-interpretation. Building detection is the first and very difficult step in building
recognition and reconstruction. It is to find buildings and separating them from the background in the presence of distractions caused
by other features such as surface markings, vegetation, shadows
and highlights. This is an instance of the well-known figure-ground
problem. The goal of automatic building detection in this paper is to roughly delineate the rooftop of the buildings that will be
verified during the recognition and reconstruction phase. The rooftop detection algorithm proposed here is based on multiple
features and proceeds in two steps: first, low-level feature extraction; second, rooftop identification. In this paper we focus on
rectangle building roof recognition. In this case, the boundaries of their rooftop are straight lines. One of the obvious facts is that
most build roofs are built of materials of limit categories, so their image colors and textures are of limited categories. Low-level
features used here are straight-line segments, image colors and image textures. In local edge detection, the vital phase of low-level
feature extraction, we introduced a novel edge detection algorithm based on EMD (The Earth Mover's Distance), which works better
than traditional ones. A general curvature concept was used in the measuring of image textures, which is invariant to rotation. And at
last with the mathematics tools of Hough Transformation and fuzzy density function we made the last decision to determine “it is
building or not”. Experiments were carried out in the Quick-bird images of Beijing, China. We were able to achieve a right detection
rate larger than 75% for those buildings that are not occluded severely.
l. INTRODUCTION
The measure of locating areas of city changes through remotely
sensed data is an important tool for many applications including
resource management, urban planning, and environmental
impact analysis, especially in the fast developing areas in China.
In fact, in China, monitoring land cover changes is currently
one priority of city land management bureau's works in the
coming years. To efficiently monitor land cover changes, we
need an efficient mean of determining where is a new building.
Till now, although many experts of diverse background have
struggled for many years to create such kind of tools, there is
still no satisfied one. We have to identify buildings in images
with tedious manual editing. In this paper, we start a new try to
develop a tool of automatic recognition of buildings from aerial
and space images without or with minimal operator interaction.
As to previous works in this field, Prof.Grun, etc, made a
excellent review of detection of buildings from aerial space
images(Grun ef al, 1995: Robert W. Carroll), some more work
is described in (Fua, 1996; Henricsson ef al, 1996; Weidner,
1996). In the last several years, more and more new kinds of
inputs, such as stereo images and range images are used. In this
paper, we still focus on the use of single-images. With little of
direct 3-D information, making use of single-images is more
difficult, but it is attractive due to the ease with which they can
be obtained. Further more, many of the processes involved in
*
616
single-image analysis are also required for multiple-image
analysis.
Our algorithm is restricted to recognize the rectangular rooftop
of the buildings in single-images, especially in color-images.
The rooftop detection algorithm proposed here is based on
multiple features and proceeds in two steps: first, low-level
feature extraction; second, rooftop identification. Low-level
features are straight-line segments and image texture. As local
edge detection is vital here; in colour images, we detect local
edge points with a novel edge detection algorithm based on
EMD (The Earth Mover's Distance), which works better than
traditional ones. After detecting local edges, straight-line
segments are found with Hough transformation. The probability
density of image curvature is used as the feature of texture in an
image patch, which is invariant to rotation. In the rooftop
identification step, we first find potential rectangles by using of
the parameters space that is got in Hough transformation. Based
on the ratio of width and height of one rectangle, the size of the
rectangle, the properties of the probability density of local
image's curvature in the rectangle and the average color of local
image in the rectangle, we compute fuzzy integrals based on the
evidences gathered to confirm whether the area belongs to a
rooftop. Our novel strategy improves the efficiency of human
Work in building recognition with few interactions.
The paper is organized as follows. First we give a relative
detail presentation of the novel edge detection algorithm based
on EMD (The Earth Mover's Distance) in section 2.Second, in
Corresponding author. This is useful to know for communication with the appropriate person in cases with more than one author.
at practical interest for many of
Internat
section
lines ant
the form
Last, in
using of
3 iul
As the
detectior
many te
Distance
images.
With f
edge det
constant
making |
statistica
Convo
because |
edge it is
as findin
the dista.
Mark A.
While
applicati.
increases
will not |
features 1
color im:
over a le
intensitie
ordering,
perceptu:
To ove
detection
remote s
Distance
fields. T
assumes 1
of an ed;
over how
between t
only the r
Beside
The first
false neg
possible i
in practic
variations
uncorrela!
without a.
distributic
binary, gr
meaningfi
generalist
2.1 A Pe
One of thi
is the dist
the distan
the groun
should coi
by psycho
