The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
284
1. INTRODUCTION
Application as city planning and environmental investigation is
crying for acquiring and updating 3D city, especially the
building, information 11] . Traditional 3D building information
was in general manually acquired from stereopair, which took
much time, manpower and money. So, many scholars have
being endeavored in researching automatic or self-automatic
extraction of building and gained much for the past 20 years [2] .
During the last few years, LIDAR, direct acquiring 3D ground
information, has become a very attractive alternative of
traditional technique for its high accuracy and lower cost.
According to different data source, there are basically two types
of LIDAR. One is incorporating LIDAR with the building’s
geometry [1,3-61 or directly depending on the layer feedback
signal [7] to extract buildings. Among which, sole application of
signal is still somewhat difficult while that incorporating with
buildings geometry is easier. The other way is combining the
LIDAR data and other data [8 ~ 10] , such as aerial image and
multispectral images, to extract buildings. But, just because of
independently acquiring of different data, it is a little difficult
either to match them and hard to recognize the borders so as to
make it somehow difficult in application.
A typical step of combining geometry features to extract
building is [5] , firstly to filter the DTM from LIDAR data, then
divide the DSM data into ground points and non-ground
points(including vegetation and building) by height difference,
finally use the segmentation method to extract buildings away
from vegetation in non-ground points group by some geometric
parameter as area threshold, gradient and surface undulation.
Although this filter-segment method having achieved much in
real application, the existing filtering methods, for instance the
morphological filtering [11 ' 13] and surface prediction [1415] , also
has obvious deficiency [13J which is especially hard to be
applied for the undulate area. No mater how to select the
segment parameter, there is a typical problem of mis-dividing
vegetation and buildings into a same classification for those
undulate regions with higher vegetation or vegetation much
close to building so as to affect the segmenting accuracy.
Since not all city landforms are flat, for which the existing
filter-segment method is hard to be applied, in this paper we
discussed the filtering, selecting geometric parameter and the
parameter sequence, and raised a new and high-accuracy
filter-segment method to extract buildings in the undulate
region. We have verified the new method by a testing area in
Virginia. Buildings were extracted out and the aerial
photographs were used to evaluate the result.
2. EXTRACTING BUIDINGS BY PROPOSED
METHOD
2.1 Filtering
The original LIDAR cloud data is discrete and irregular, which
needs effectively post-processing. It is firstly to classify
effectively and exactly the ground points from non-ground
points, which we called filtering.
1) Extracting DTM
In order to extract exact DTM in fluctuant regions, this paper
proposed a new filtering method called contour-based surface
estimation. For such a truth that vegetation and building points
are much higher than the ground points around them in some
areas, and its contour lines (hereinafter referred to as Object
Counter Line, OCL) are obviously close (1 and 2 in Fig. 1),
while contour line formed by ground points (hereinafter
referred to as Ground Counter Line, GCL) are always unclose
(3 in Fig.l). Thus we can easily differentiate OCL from GCL
according to their closing status. However, the closed OCL at
image edge is easy to be manually cut off while cutting images
(4 in Fig.l). In this case, we can use the parameter as the
distance from start to end of the contour line and the curve
length to distinguish them.
From the already out extracted GCL, we can get DTM which is
not exact enough at this stage. It is mainly because few GCLs
of downfold and hill were also removed while judging the
closing of contour lines thus leads to the phenomena of hill
cutting or downfold refilling in DTM. To resume the ground
points, we can use the iterative approaching method to refine it
as follows:
3. Take the DTM that interpolated from the GCL as the
initial DTM,
4. Compare the elevation between DSM points and the
initial DTM, take any DSM with the point elevation
difference less than the threshold Ah i(e.g. A/z^O.lm)
as a new DTM data; contrarily the new DTM of this point
shall be set as blank,
5. Count the points increment ( An ) with elevation
difference less than the threshold, if An less than 100,
interpolate those points without data in new DTM from
all the points with data, get the final DTM and output.
Otherwise execute the next step,
6. Interpolate those points without data in new DTM from
all the points with data, and replace the initial DTM by
the new DTM, return (2).
2) Normalized DSM
So-called normalized DSM is actually DSM-DTM. Since
buildings are higher than ground, if we set a elevation
difference threshold Ah 2 (e.g. a general Ah 2 =2.5m), using
this we can remove the lower objects as ground points and cars
and get a initial building segmentation surface including
vegetation information but excluding ground information.
Because it is impossible to get a completely accurate DTM
extraction, the initial building segmentation surface gained only
by elevation difference contains possibly some ground points,
moreover, the geometrical parameter (e.g. acreage and
quadratic gradient) of those points is quite similar to those of
the building roof, which is hard to be removed in the following
process hence to affect the precision of extracting buildings.
Necessary measures have to be adopted to remove them.
Considering the elevation of buildings and ground have the
following features: building surface is relative regular but the
elevation difference at the edge changes evidently with strong
Fig.l Contoured aerial photograph