DATA
metric and remote
nd this opens the
abilities of image
c DSM generation
o-fine hierarchical
SMs are generated
10to geometrically
/ith multi-image or
> accuracy tests are
measurements and
e demonstrate with
zes (Gruen, Zhang,
1 extended and has
s as well. We will
for our procedure.
s. We will give
SI, IKONOS and
TONS
nage matching has
\ wide variety of
automatic DEM
vile commercially
tric workstations.
trategies used may
formance and the
major systems and
ers does by far not
irements (Gruen et
3M generation are
tions and others
uence the matching
lerived from 5 m
e IKONOS images
n 5 cm pixelsize SI
h-resolution aerial
lusions, the surface
d trees, large areas
1s; etc.
provides for new
ching:
ely 12 bit) 1mages,
tches even in dark
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
e |t has the ability to provide multiple images with multiple
channels. Thus it enables the multi-image matching approach,
which leads to a reduction of problems caused by occlusions,
multiple solutions, surface discontinuities and results in higher
measurement accuracy through the intersection of more than two
image rays.
e It has the ability to provide for relatively precise orientation
elements that can be used to enforce geometric constraints and
restrict the search space along quasi-epipolar lines.
eThe nearly parallel projection in along-track direction causes
less occlusion on the nadir-viewing images.
r Images and Orientation Data |
Image Pre-processing & Image
Pyramid Generation
L |
i Geometrically
Constrained
Grid Point
Matching
Candidate Search,
Adaptive Matching
Parameter
Determination
Feature Point
Matching
| Edge Matching
v
DSM (intermediate)
Combination of the feature points, grid points and
edges
Modified Multi-image Geometrically
Constrained Matching (MPGC)
Final DSM
Figure 1: Workflow of our image matching procedure
Among the known matching techniques and algorithms, area-
based (ABM) and feature-based matching (FBM) are the two
main ones applied to automatic DSM generation in general.
ABM and FBM have both advantages and disadvantages with
respect to the problems presented above. The key to successful
matching is an appropriate matching strategy, making use of all
available and explicit knowledge concerning sensor model,
network structure and image content.
Our matching approach is a hybrid method that combines ABM
and FBM. It aims to generate DSMs by attacking the problems
(a)-(f) mentioned above. It uses a coarse-to-fine hierarchical
solution with a combination of several image matching
algorithms and automatic quality control. Figure 1 shows the
workflow of our matching procedure. After the image pre-
processing and production of the image pyramid, the matches of
three kinds of features, i.e. feature points, grid points and edges
on the original images are finally found progressively starting
from the low-density features on the images with low resolution.
A triangular irregular network (TIN) based DSM is constructed
from the matched features on each level of the pyramid, which in
turn is used in the subsequent pyramid level for the
approximations and adaptive computation of the matching
parameters. Finally the modified MPGC matching is used to
achieve more precise matches for all the matched features and
identify some false matches. In the MPGC procedure, multiple
strip image data can be introduced and combined. More details
of our approach will be provided in the next paragraph.
3. THE MATCHING APPROACH
3.1 Image Preprocessing
In order to reduce the effects of radiometric problems like strong
bright and dark regions and to optimise the images for
subsequent feature extraction and image matching, a pre-
processing method, which combines an edge-preserving
smoothing filter and the Wallis filter. First, the edge preserving
smoothing filter proposed by Saint-Marc et al., 1991 was applied
fo reduce the noise, while sharpening edges and preserving even
fine detail such as corners and line endpoints. Next, the Wallis
filter, which strongly enhances already existing texture patterns,
IS applied. The examples of Figure 2 indicates that even in
129
shadow and "homogeneous" areas the image content is
enhanced.
The image pyramid starts from the original resolution images.
Each pyramid level is generated by multiplying a generation
kernel and reduces the resolution by factor 3. The pyramid level
number is a pre-defined value that is either a user-input or can be
determined according to the height range of the imaging area.
Figure 2: SI image before and after pre-processing
3.2 Feature Point Matching
We use the Foerstner interest operator to extract well-defined
feature points that are suitable for image matching. Firstly the
reference image is divided into small image patches (the nadir
viewing SI or satellite image is selected as reference). Only one
feature point will be extracted in each image patch. The density
of the feature points can be controlled by the size ofthe patches.
The determination of the correspondences of the given points on
the search images is carried out using the geometrically
constrained cross-correlation method (see Gruen, Zhang, 2003).
The matching candidates are computed by cross-correlation with
a set of adaptively determined parameters like the image window
size w,, the threshold of the correlation coefficient c, and the
search distance. The approximate DSM that is derived from the
higher-level of the image pyramid is used to estimate these
parameters.
We incorporate the method proposed by Kanade & Okutomi,
1994 to select an optimal window size by evaluating the image
content and the disparity within the matching window. As a
result, in flat areas with small image intensity variations, the
window size w, increases and in areas of large terrain elevation
variations it decreases. The threshold of the correlation
coefficient c, should also vary according to the roughness of the
terrain. We set a larger value in flat areas and smaller value in
rough terrain areas. The search window size depends on the
terrain elevation variation in a small neighborhood of the given
point and on the image geometry. In relatively flat areas the size
of the search window decreases and vice versa.
By adaptive selection of these parameters, we can both reduce
the processing time and the probability for multiple matching
candidates. The number of matching candidates can be further
reduced by introducing a third image. For every candidate, its
position on the third image can be predicted using the image
orientation elements. If the correlation coefficient between the
reference and the third image is lower than the threshold, this
matching candidate will be dropped. However, we cannot
completely avoid the ambiguity problem due to reasons like
repetitive patterns. Our procedure takes n (X 5) matching
candidates with correlation coefficient values above the
threshold c;.
As a result, for each feature point on the reference image several
matching candidates can be computed. The correct match is
determined by analysing the following quality measures
sequentially:
a) The correct match should have a clear and sharp maximal
correlation coefficient. If there are more than one candidates and
the value of the first correlation coefficient peak is more than two
times of that of the secondary peak, the candidate that has the
largest correlation coefficient value will be considered the correct
match.
b) Using the same matching parameters, the feature point can be
back-matched from the search images to the reference image. If
the difference between this two-way matching is less than one
pixel, the candidate is assumed to be the correct match.
c) Under the condition that the feature point appears on more
than two images, the residuals of the photogrammetric forward
intersection should be less than 2-3 times the standard deviations
of image coordinates of the triangulation adjustment.
