International Archive
derived digital orthoimages the creation of 'GIS'-proof DCMs
will be discussed.
2. BUILDING EXTRACTION - STATE OF THE ART
Airborne laser scanning has become one of the most important
acquisition techniques for surface reconstruction. By collecting
a huge amount of points (3D locations) with a very high
resolution and accuracy they provide an optimal input for
building extraction algorithms. Several approaches have been
presented, among them one by Rottensteiner's (2004), which
can be divided into three major steps. The first one is to
hierarchically apply a robust interpolation using a skew error
distribution function in order to separate terrain points from
points that lie not on the ground. The outcome is a DTM
(Digital Terrain Model). The second step is to distinguish
between building points and off-terrain points (e.g. vegetation).
This is achieved by applying an analysis of height differences
between the DSM and the previously computed DTM. Finally,
the third step comprises the creation of polyhedral building
models.
A different method for automated building reconstruction is
described in Peternell and Steiner (2003). Assuming that a
dense point cloud is given, they try to determine all planar
regions of a building. Firstly, local regression planes for all
points are computed. The planar faces of the building possess
local planes of regression (local planar fits) that are close to
each other. The finding of the faces is achieved by
transforming the local planes of regression into a parameter
system where a segmentation is carried out. A drawback is the
fact, that additional input information (digital map) is needed
for defining the area of an object.
Both approaches need a dense and highly accurate point cloud
in order to be able to derive and analyse the roof faces of
buildings. They aim at a rather large scale reconstruction.
Unfortunately these algorithms cannot be adapted to point
clouds created by spaceborne or line-scanning systems since the
geometric resolution as well as the accuracy and completeness
is usually insufficient.
3. CHANGE DETECTION / UPDATING - STATE OF
THE ART
Most of the algorithms for updating check the existing
buildings in the data base for changes. They are not able to find
new objects at locations with no entry in the old data base.
The ATOMI project (Eidenbenz et al., 2000) was used to
update vector data of building roof outlines and increase their
accuracy by applying image analysis techniques on aerial
images. For finding new buildings Niederoest (2000) tried to
localize extracted DSM blobs or apply a multispectral
classification. In order to extract DSM blobs that really
correspond to buildings we have to assume that there exists no
big and high vegetation and that the DSM resolution and
quality is good enough to be able to separate buildings from
each other or from trees. The idea of applying a multispectral
cassification in order to determine new buildings is of course
only relevant if multispectral data is available, which is
unfortunately not the case for high-resolution satellite imagery
most of the time.
Knudsen's (2003) strategy for updating digital map databases
makes also use of pixel-based classification. The algorithm
uses vector and spectral data as input to an unsupervised
spectral classification method that controls a subsequent
Mahalanobis classification step. Due to good geometric
s of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
resolution and the potential of using RGB images, changes may
be traced and even new buildings detected, although again due
to the pixel-based approach obtaining topologic information of
the new situation becomes cumbersome.
4. WORKFLOW
The proposed workflow is divided into two major parts. The
first one is data preparation and comprises all steps from data
acquisition to the derived DSM. There are many off-the-shelf
software packages available for solving this task. The second
part deals with the actual creation of the DCM including
building recognition and extraction. Figure 1 gives an overview
of the proposed workflow.
Data Acquisition
L Satellite / 3-line sensor
T E
Y
Stereo Matching Multi-Image (Ray) Matching]
Digital Surface Model rer rie Orthophoto
|
Shih ERE HU, t A |
Digital Terrain Model Normalized Surface Model |
ces |
|
Thresholding / Masking |
| |
niue |
Building Recognition |
ER, M e s pu E
Old (City) Database | Building Extraction pere
dq
Updating apres Digital City Model
Visualization
Figure 1. Proposed workflow
|
|
* : |
Orientation (RPCs) r Orientation (IMU/GPS) |
"esp — —— 575 —— |
|
|
|
l
4.1 Data Preparation
High resolution satellites deliver overlapping image scenes in
form of stereo pairs, whereas multi-line scanners generate one
image strip for each sensor line, where each strip is associated
with a different viewing angle. The orientation of multi-line
sensor data can be automated to a very high degree.
Vozikis et al. (2003) have shown that images from the new
high-resolution satellite sensors (IKONOS and Quickbird)
cannot be oriented by applying the common procedures used in
acrial photogrammetry, because, firstly, most distributors do
not provide any calibration information (interior orientation
parameters) and secondly, due to the extremely narrow field of
view (bad intersection quality of the rays), the well-known
collinearity equation would fail. Diverse orientation models
have been proposed, but the fastest way is employing the so-
called RPCs (Rational Polynomial Coefficients), which are
usually provided together with the images. These coefficients
describe a relation between object and image space and should
give the end-user the possibility to start immediately with
image exploitation. Unfortunately the RPCs are affected by a
constant shift and hence are not very accurate, SO they have to
be refined by using two or three GCPs (Ground Control Points)
(Hanley et al., 2001).
For the airborne line scanner image strips, the orientation
procedure is a little bit different. Together with the image data
1034
In
ac
ar
Ve
CA
ac
me
Fi
ort
bu
At
hai
aut
nec
201
wh
ori
pro
4.2
The
Mo
and
Sca
Her
pro
The
Mo
DSI
ele
The
appi
a sl
fully
of tl
in v
whe
poin
the |
digit
DT^
The
inter
Unti
builc
had t
Rott:
a cet
wher
even
'Non
