matching window and for different levels of the pyramid
images. If the approximate range of elevations in the
overlapping area is already known, it is easy to derive a
searching range for different cases. Otherwise approximate
geometric transformation parameters can be used to derive it.
The searching range should decrease with increasing the
level of pyramid images when we consider image resolution.
On the other hand, when we consider the matching strategy
from the top level down to the original image, i.e., acquired
approximations are getting better from the top level down,
does the searching range decrease with decreasing the level.
Generally, we define the searching range as large as possible
on the top level of the pyramid images to avoid the
possibility of losing matching.
3. DISCUSSION
In this section some characteristics of the RG-DW
matching scheme are discussed.
3.1 Simplicity
As described above, the RG-DW scheme combines the
advantages of very simple elementary tools. Without
introducing any complicated concepts and algorithms, and
without any a priori knowledge about images and object
space, the RG-DW can be easily programmed and applied to
match satellite images and aerial photos.
3.2 Reliability
For each anchor point a total of at least 12 windows of
different sizes are used to do the matchin g. They are
compared to each other, so that the percentage of wrong or
lost matches drops dramatically compared to the case where
only a single window is used to do the matching.
3.3 Accuracy
Many experiments with the RG-DW show that the
matching accuracy is very satisfactory, therefore the results
can certainly be used as very good approximations of a
successive least squares matching to get subpixel accuracy.
3.4 Independence
As mentioned before, the RG-DW does not rely on any a
priori knowledge about image orientation or information
about object space. The only thing to do before performing
matching is to very roughly determine the four corner points
at the overlapping area of the stereopair of images. The
whole matching process can be completed independent of
other processes such as image orientation. In other words,
once a stereopair of digital images is on hand, matching with
MATCH can be implemented. On the other hand,
approximations for all points to be matched can be derived
from the first geometric transformation parameters, i.e., each
point can be processed independently.
3.5 Parallelism
It is clear that this characteristic directly comes from the
independence of the RG-DW noted above. Because the
required initial values for cross-correlation of each point are
directly derived from the geometric transformation
parameters between the two images, it is very convenient to
perform a parallel processing, i.e., all points to be matched
can be derived at the same time, if array processors or
parallel computers are available.
On the other hand due to this parallelism, systematic
matching errors are greatly eliminated.
4. RESULTS
The software package MATCH based on the RG-DW is
the kernel of a new commercial digital photogrammetry
module, the Digital Ortho Module of ERDAS Inc.. It has
been applied for both SPOT images and aerial photos. The
success rate of matching results (SRMR) is quite high. For
two pairs of full SPOT (1A) success matching was
performed at different grid intervals. For the case with an
815
interval of 240 pixels, 484 (22 x 22) anchor points were
matched, which are homogeneously distributed over the
whole overlapping area. The point were visually checked on
the screen of a Sunsparc 1 workstation with full image
resolution. The SRMRs of both pairs are 96 percent. Most
wrongly matched points fell into clouds or empty areas
without information. Figures 3 and 4 show one of the pairs.
Figure 3 is the pair of original images with 6000 x 6000
pixels and figure 4 shows the anchor point matching results.
The SRMR for densification is similar as the anchor points
according to the visual color examination method mentioned
above. If those anchor points with lower reliability indices
values had been interactively corrected, the SRMR of
successive matching would increase. Figure 5 shows
densification matching points in a small area (512 x 512
pixels) of figure 3. It can be seen that even though there
exists quite a big geometric distortion, the matching results
with the RG-DW are still good. Figure 6 is a DEM created
from the matching results of figure 3. Figure 7 is a
perspective view derived from figure 6.
Figures 8 to 12 are results of aerial photos digitized with
a pixel size of 100 microns. Figure 8 shows a pair of
original images and some anchor points. Figure 9 and 10
display two parts of densification matching results (512 x
512 pixels). Figure 11 is the DEM derived from the
densification matching results of the overlapping area.
Figure 12 is a perspective view of the DEM overlaid with the
orthophoto. Figure 13 shows matching results from aerial
photos taken by a CCD camera installed in a helicopter.
It should be noted that all results displayed here are direct
results without any correction by the human operator, so
some inconsistent situations may be found in the figures.
For visibility, the sizes of crosses representing matched
points in the images were enlarged.
Although the authors did not optimize the program, it
is acceptable for current use, for example, the run time of
MATCH for the full scene of figure 3 with intervals of five
pixels on a Sunsparc 2 was about 24 hours.
5. OUTLOOK
The RG-DW scheme with its simplicity, reliability and
parallelism, and its practical application shows that it is a
promising image matching technique. Further investigations
will be concentrated on the following aspects:
1. Postprocessing - make use of some existing
conditions to automatically detect and delete Wrong
matches.
2. Investigate and design parallel algorithms with the
RG-DW.
REFERENCES
Ackermann, F., Hahn,M., 1991: Image Pyramid for Digital
Photogrammetry, Proceedings of the 43rd
Photogrammetric Week, Heft 15, Stuttgart
Lue, Y.and K. Novak, 1991: Recursive Grid-Dynamic
Window matching for automatic DEM generation,
Proceedings of 1991 GIS/LIS and ACSM-ASPRS Fall
Convention, pp. A254-260, Atlanta
ERDAS Inc., 1992: Field Guide of ERDAS
Lue, Y.and K. Novak, 1992: An Operational Image
Matching Package «MATCH», Proceedings of 1992
ACSM-ASPRS Annual Convention, Washington D.C.
Novak, K., 1992: Application of Digital Cameras and GPS
for Aerial Photogrammetric Mapping, Presented paper
of Commission I of XVII ISPRS Congress,
Washington D.C.