calibration edge is on the left when the camera is viewed from
the back.
For example, against the order of (1 2 3 4 5 6 7 8) for the
calibration edge located on the left as shown in Figure 1, the
orderof (43128756),(21436587)and(3421786
5) will be automatically used for the calibration edge to the
top, right and bottom respectively, as chosen by the user.
Quality control
At the end of the fiducial searching process for each frame, a set of
transformation parameters is derived based on the coordinates of
fiducials in two different systems. If the R.M.S.E. is over a certain
threshold, the software will warn user of the results and also
record it in the A/O report for later post-checking when executed
in batch mode. During the processing if the correlation coefficient
of the LSM for any fiducial is lower than the threshold, it will be
automatically rejected and logged in the 4JO report.
1.2 Results
Before the beta version of the AIO was offered to Vision's
users in early Spring of 1995, a total of over 400 digitized b/w
and color frames with different image scales, scanned by
different commercial scanners with different formats and
different scanning resolutions ranging from 15 to 30 microns
per pixel, were tested. One hundred percent success rate of
fiducial recognition with a 0.1 subpixel positioning accuracy
and a RMSE of 0.325 pixels for coordinate system
transformation was achieved. The performce time varied from
5 to 9 second per frame with eight fiducials, which is largely
depending on the image and scanning quality (Lue, 1995).
Recently to respond to OEEPE's workshop, the author ran
AIO on a SGI Indigo2 with R4400/200MHz workstation for the
OEEPE FORSSA data set of 28 frames taken by camera Wild
RC 20 and scanned by Zeiss PS1 scanner with scanning
resolution of 30 microns per pixel. The 4/O running time for
all 28 frames was 4°46’, 1. e. less than ten seconds per frame
with eight fiducials. The correlation coefficients for all 224
fiducials are over 0.9 and the R.M.S.E for the transformation
was 0.17 pixels or 5.21 microns.
As further proof, Vision customers have reported satisfactory
results with 4/O on thousands of different digital frames.
2. FULLY OPERATIONAL AUTOMATIC TIE
POINT SELECTION (ATPS)
The topic of automatic aerial triangulation (447) has been
discussed a lot recently. It is well known that the aerial
triangulation consists of two major phases: mensuration and
calculation. The latter one has been successfully solved in
1970's and achieved its very high accuracy to meet all kinds of
application for several decades. The focus of AAT for DP has
naturally fallen onto the former one, which corresponds to the
tie point selection, transfer and mensuration of the image
coordinates. That was the motivation to develop AJO first and
then ATPS, since much of AIO is a prerequisite of AAT. If and
only if we have AIO and ATPS together with AT we could say
we have been reaching the AAT.
On the other hand, there are many factors which determine the
final accuracy of the A7, such as flight quality including the
situation of overlapping, picture taking, quality of scanning,
480
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
quality of the ground control points, the capabilities of the AT
software, as well as the configuration, accuracy, reliability and
number of tie points. Actually, only the last four factors are
derived directly from ATPS and require our attention at this
stage to implement the AAT.
According to Vision's customers - productive photogrammetry
companies - the A7 results derived from the semi-automatic
tie points selection and measuring of the SoftPlotter"M met all
standard requirements of conventional ones. For this reason, to
verify the A7PS results without involvement of other factors,
which might affect the final accuracy but out of control of
ATPS software, we simply chose the visual check on the
computer screen to get the statistics of success/failure instead
of through the A7 results.
A similar philosophy and strategies - like multi-level image
matching with a dynamic window size, spiral searching and
LSM etc. - are still used for ATPS, though feature extraction, a
variable threshold for different level of pyramid images and a
more sophisticated organization with strategies distinguished
it from AIO.
2.1 Basic Concepts And Technical Strategies
Basic tools
The basic tools used in ATPS are: 3-4 levels of pyramid
images, point-like feature extraction, multi-scale matching,
spiral searching strategies, LSM.
Four basic factors
As mentioned above, configuration, accuracy, reliability and
number of tie points are only factors controled by ATPS which
will directly affect the final AT results. With this
consideration, the measures adopted by ATPS in order to meet
AT 's requirements for these four aspects are:
Configuration: standard position (Gruber points) with six more
additional patterns (Figure 4)
Accuracy: using LSM at the final step to achieve high
accuracy
Reliability: using the multiple criteria control to reduce the
rate of wrong matching
Number: 3 x 3 points per frame at least (see Figure 4). At
each position single/cluster point is possible.
3-4 pyramid image:
Similar as AIO, one of the most important strategies to make
the ATPS feasible is the use of pyramid images throughout the
processing. The levels of pyramids are separated by a scale
factor of 4. The total levels for A7PS are normally about 3-4,
depending on the scanning resolution (the finer resolution, the
more levels).
Automatic overlapping range prediction
The ATPS starts from an automatic prediction of percentage of
overlapping range between the immediately adjacent images
from the highest pyramid image to determine where to put the
selected tie points and from how many images (two/three/five
- stripwise or four/nine/fifteen - blockwise) and which images
to do matching to find conjugate points. The horizontal and
vertical parallaxes obtained in this stage are also be used to
partly guide the subsequent searching processing. The
percent
betwee
resolut:
Searchi
of the
always
proces:
the py
levels :
LSM 1s
Point-i
A poin
central
the nex
locatio
The m
level v
search
show
matchi
additio
on the
out ag
interes
versior
The re
certain
levels :
on the
feature
conceri
will w
any fea
Option
The e
approx
parame
the siz
feature
point c
parame
autom:
In eac
needec
feature
the int
and ba
patch
using :
of feat
best fe
It sh
photog
points,
tie poi
even t
point 1
enougl
differe
choose
