'. Istanbul 2004
resolution)
ited points over
- no texture. The
mputed by using
d to the feature
thus many grid
ture. The search
sibilities to yield
ratching method
s the candidates
from their local
ains the highest
ny’s “coherence
rate this idea in
on by relaxation
sular grid in the
1e search images
ature points and
here that all the
classes: Points
candidates, and
st case, they are
nd, they serve as
‘or the last case,
yith a very small
within the search
ines in the TIN
ocal smoothness
key point of this
bint matching, is
thness constraint
traint, areas with
> bridged over,
ly over the area.
an be preserved
ss the edges. For
3).
M is constructed
yramid, which in
level for the
f the matching
aklines such that
tains the original
n. The surface
d and preserved.
e pyramid can be
d on a “region
uses the already
; and matches the
cal neighborhood
88). This method
an be treated as
on image pyramid
pine terrain, some
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
manually measured points can also be introduced as seed points.
It will give better approximations for the matching.
3.6 Refined Matching Based on the Modified MPGC
MPGC (Multi-Photo Geometrically Constrained Matching) was
developed by Gruen, 1985, and is described in detail in
Baltsavias, 1991. It combines least squares matching and
geometric constraints formulated either in image or in object
space and permits a simultaneous determination of pixel and
object coordinates. Any number of images (more than two) can
be used simultaneously. The achieved accuracy is in the sub-
pixel range.
Our modified algorithm is an extension of the standard MPGC. It
integrates the geometric constraints derived from the Linear
Array sensor models. The geometric constraints force the
matching to search for a conjugate point only along a small band
around the epipolar curve and reduce the possibility of false
E A | 4i ua =
Figure 4: MPGC matching with multi-strip SI images.
Top: Images of a strip acquired from west to east
Bottom: Images of the cross-strip
The modified MPGC is used to refine the matched features in
order to achieve sub-pixel accuracy. The DSM derived from the
approaches (3.2)-(3.4) provides quite good approximations for
the MPGC procedure and increases the convergence rate. The
initial values of the shaping parameters in MPGC can be
predetermined by using the image geometry and the derived
DSM data. The corresponding rays of the four corners of the
matching window in the reference image are intersected with the
derived DSM and their object coordinates can be determined.
Through imaging geometry, the corresponding image
coordinates in the search images can also be determined. The
initial values of the shaping parameters can easily be computed
from these four corner points and their correspondences. By this
way, features can be matched with multiple images or even with
multiple image strips that have different flight directions and
image scales. Figure 4 gives an example for the case of cross-
strips.
For edge matching the parameters of a spline function of the 3D
object edge are directly estimated together with the matching
parameters of edges in the image spaces of multiple images.
Some points, especially those grid points in non-texture or little
texture areas, will fail in MPGC matching. These points are also
kept but they are assigned low reliability indicator values.
4. EXPERIMENTAL RESULTS
In order to evaluate the performance of our approach for DSM
generation it has been applied to different areas with varying
textures, terrain types and image scales. In the following we will
report about DSM accuracy test results with SI, IKONOS and
SPOTS HRS images.
4.1 SI Image Dataset, GSI area, Japan
In Japan's GSI (Geographical Survey Institute) test area, both SI
images and aerial photos are available. The evaluation is based
on the comparison between the manually measured DSM from
aerial photos and the automatically extracted DSM from the SI
images.
131
The GSI test area is roughly 650 x 2500 m”. It is relatively flat
with natural and artificial objects. There are a lot of small
geomorphological features, but also small discontinuities like
cars, isolated trees and large discontinuities and occlusions due
to buildings. For the success of a matcher it is very important
how it can handle local discontinuities (e.g. buildings or other
man-made objects, vertical cliffs, etc.).
Figure 5 shows two image windows from the nadir-viewing SI
images. We have 48 control points. They are signalized marks
on the ground or on the top of buildings. They appear both in the
SI and aerial photos.
Two stereo pairs of color aerial images of 1:8000 image scale,
acquired with a film camera of 153 mm focal length, have been
used for manual collection of reference data on an Analytical
Plotter. The RMSE of the exterior orientation is reported as 5 cm
in planimetry and 3 cm in height. Manually measured points,
distributed at about 50 cm distance in both X and Y directions,
are used as reference data.
Assuming a measurement accuracy of 0.1% of the flying height
(best case situation for natural points) we can expect an accuracy
of the reference data of 0.12 m as a best case scenario.
x
Figure 5: Region (left) of man-made objects. Region (right) of bare
terrain area (including some sparse trees). The image patches are
from the nadir SI-100 images with the ground resolution of 5.6 cm.
Figure 6: 3D visualization of the shaded DSM of the GSI test area.
Table 1: DSM accuracy evaluation.
Manually measurement minus automatically extracted DSM
Area | RMSE | Mean % % % % %
(m) (m) 0.0-10m | 1.0-20m | 2.0-30m | 3.040m | >4.0m
(1) 0.28 0.04 98.65 0.91 0.13 0.09 0.03
(2) 1.01 -0.16 74.70 12.15 7.32 4,84 2.36
|
The processed SI data include three panchromatic images with a
footprint of ca. 5.6 cm. As a result of triangulation, 2.8 cm and
5.0 cm absolute accuracy in planimetry and height (as computed
from checkpoints) were obtained.
For analysis of the matching accuracy, we divided the reference
data into two classes, i.e. (1) the bare terrain area with some
sparse trees and small artificial objects (including the large
parking areas); (2) the area with man-made objects and trees.
Our analysis has been performed for these two classes
separately.
A very dense raster DSM with 15 em interval was interpolated
from the automatically matched point cloud and edges. The
points with low reliability values were given a small weight in
the interpolation procedure. Figure 6 shows the 3D visualization
of the extracted DSM.
Comparing the reference points with the raster DSM in two
different areas (1) and (2) leads to the results shown in Table 1.
The bare terrain area shows much better results, but it still
suffers from problems like multi-temporal differences between
