: Automatic 
ion, transfer 
nternational. 
lly available 
for fiducial 
'cessed with 
ht fiducials. 
>s. The beta 
m 1:3400 to 
'sert, rivers, 
as about 85 
on 16 data 
resolutions 
| area with 
ert, rivers, 
ed in the 
1e last part 
ATIC 
) 
nts of all 
feature or 
era system 
ocessing a 
gitizing. A 
d to derive 
versa. The 
eriving the 
image and 
ic tasks of 
has been. 
orientation 
Interior orientation in most existing systems require at least 
the first one or two fiducials to be measured manually before 
the remaining fiducials coordinates can be determined by 
various semi-automatic methods. On the contrary, AJO of 
SoftPlotter!M ^ implements interior orientation fully 
automatically without the need for any approximations or 
intervention by a human operator. 
The approach used for A/O 1s a revised version of the RG-DW 
matching scheme, developed for Digital Ortho Module of 
ERDAS 7.5 for DEM generation (Lue, 1991, 1992), 
augmented with the successive LSM to yield with very high 
accuracy. For simplicity, we outline the methods in the 
following part. The reader is encouraged to refer to the 
literature for the detail. 
1.1 Basic Concepts And Technical Strategies 
Basic tools 
The basic tools used in AIO are: three levels of pyramid 
images, template matching, spiral searching strategies, least 
squares matching (LSM). 
Take full use of a priori knowledge 
Compared with other image matching problems, the searching 
for the camera fiducials is simpler, because the fiducial has a 
known shape and location on the digital image. This kind of a 
priori knowledge can be easily exploited to simplify searching 
and processing. 
In general, the fiducials are normally located on the corners 
and/or on the edges of a film as shown in Figure 1, and 
different cameras have their own fiducials with different 
shapes. To perform the template matching for different 
fiducials a set of fiducial templates is needed. An easily 
extensible database containing templates of fiducials for 
different aerial cameras has been established for AIO through 
scanning of fiducials with a very fine resolution, as shown in 
Figure 2. Using the fine resolution allows the templates to be 
better resampled to match the scanned pixel size of any input 
image. 
Clearly, it is unnecessary to work on an entire digital image to 
perform the fiducial template matching. As mentioned above, 
the a priori knowledge about the fiducials positions allows for 
a quicker searching of a patch of pixels only surrounding the 
predicted fiducial location, eliminating the need to generate 
pyramid images for the entire frame of original image for AIO 
use. 
Template matchig on three pyramids and L$M work 
together 
It is essential to achieve a good approximation prior to LSM. 
The template matching is less sensitive to poor initial 
approximations than is LSM, while LSM typically provides 
better final results than template matching. Therefore more 
attention was provided to developing strategies to assure 
reliable template matching to provide better initial value to 
guarantee the desired LSM results. 
For the first fiducial, a small patch, say 512 by 512 pixels, is 
read from the original digital image. In order to get a higher 
efficiency for fiducial searching with less effort, three levels of 
pyramid images (Figure 3) and the original image for each 
small patch are used throughout the template matching 
processing, e.g. in case of 25 microns of scanning resolution 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
for an original image the resolution for its three pyramid 
images will be 100, 400 and 1600 microns respectively. More 
effort and a relatively wider search range for the first fiducial 
are normally required due to the limited information to predict 
its position. 
The matching starts at the lowest pyramid resolution level, 
and the solution obtained is then used as a starting point for 
the next level's matching. A set of dynamic correlation 
coefficient thresholds and dynamic window sizes for the 
template matching are adopted. In general, a lower threshold 
for a higher level and vice versa to avoid a possible wrong 
recognition for a lower level or lost matching for a higher level 
to ensure a higher success rate of recognition. A spiral 
searching strategy (Lue, 1991, 1992) is used to locate each 
consecutively smaller patch and different searching ranges 
during the spiral searching process are accordingly used 
within each pyramid level. 
Successful location and mensuration of the first fiducial allows 
for the computation of a translation bias so that a smaller 
search range can be used for the second fiducial to gradualy 
reduce the effort. Its location can be roughly predicted using 
the computed translation bias. A satisfactory result can 
therefore be reached using a smaller patch size, say 256 by 
256, vice 512 by 512, when searching for it. The fiducial 
diagonally opposite to the first one is always treated as the 
second one for geometry consideration for later use. 
Once the first two fiducials are located successfully, initial 
transformation, scale and rotation parameters between the 
scanned and camera systems can be roughly calculated. This 
transformation is used to predict the locations of all other 
fiducials. As a result, the remainder of the fiducials can be 
located with an even smaller search patch size, say 128 by 
128, as the AIO proceeds. Once all fiducials are well located 
the final transformation parameters are calculated again and 
saved for subsequent use. 
Some practical aspects 
If the scanned fiducials are located too close to or too far away 
from the border of the scanned images, which sometimes 
happens, the search for the initial fiducial may fail. To survive 
such situations the algorithm sequentially attempts to locate 
other fiducials as the first one. If it fails again a larger patch 
size will be used in searching for the first fiducial. Then the 
whole processing is repeated. 
Sometimes one or two fiducials might be missing or obscured. 
This should not significantly affect the final results, because 
the remaining fiducials which span the image provide more 
than adequate observations for solving the interior orientation 
parameters. However, at least three, or for safer, four fiducials 
are required to achieve good quality of the transformation 
results. 
In the case of the flight direction is different from the scanning 
direction the search and measuration will yield incorrect 
results. The difference in fiducial ordering due to scan 
direction must be taken into account. The software simply 
provides a way to let the user identify the orientation of the 
imagery with respect to the flight direction by identifying a 
calibration edge as an input parameter if the direction is 
different from the default. Figure 1 presents the fiducial 
numbering sequence from a United States Geological Survey's 
(USGS) aerial mapping camera calibration report. The default