s the refer-
all matched
hanging the
ast squares
ant over the
nmetry may
‘ both solu-
because a
| from mask
rtain differ-
n the image
on the esti-
aces (which
0). In addi-
istency with
ternal mea-
iis may lead
tches
cessed suc-
2d. Now the
ich aims at
lly matched
Je to image
matching of
r more than
ns in figure
s continued
ented or all
sed so far.
le to take a
mplate con-
t this leads
d with each
template is the hypothesis that the template is a reason-
able model for the target. Instead of evaluating the match-
ing with one template the algorithm has to choose the best
match among all template matches. With the best match
the localization and identification of a certain signal is given.
For the evaluation we use the results of the self-diagnosis
module of the matching algorithm which answers
- with 0 if matching of a template with an image is done per-
fect,
- with 1 if matching is successful but with lower correlation,
- with 2 if the transfer into an image has to go the indirect
way (step 2),
- with 100 if matching fails.
The sum of this values is calculated over all matches of a
template with the images of a signalized point and is taken
as the description length for evaluating the matching with
a specific template. The template which yields the minimal
description length is considered to give the best measure-
ments.
In practice we have to take into account that signalization
for an image flight is done with relatively small signals.
Scanning of the photographs with a moderate pixel size,
e.g. with 15 um, leads to some pixels diameter of the im-
aged target. Most commonly used are round and square
targets or crosses.
Overview Detail
ue
Figure 3: Examples of signalized points. The left column
shows overviews, the right columns detailed views (21 x
21 pixels) of some signals.
Further, the varying background problem is taken into ac-
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
count within the developed algorithm. In practice the sig-
nals are located in natural terrain which makes proper
modelling difficult. Some examples taken from the flight
experiment used in this investigation are plotted in figure
3. Elimination of inhomogeneous background within the
matching process is managed by weighted least squares.
The weights can be derived from the template image, for
example, by creating a weighting mask proportional to the
gradients of the template. Another possibility is to use a
circular weighting function which steeply descend outside
a certain radius. In the experiments the circular weighting
function with a binary inside - outside decision is used.
3. EXPERIMENTAL INVESTIGATIONS
In the experimental investigations we process the image
data of a test flight experiment which was carried out in
1995. The images have been taken with a RMK TOP 15,
the photo scale is 1 : 13 000, the flying height above ground
2000 m. The block covers an area of 4.7 km x 7.2 km.
Three strips are flown in east-west direction with 7 pho-
tographs in each strip and an overlap of of 60 ?6 within and
across the strips. Another three strips are taken in north-
south direction with 5 photographs per strip and an overlap
of p = 60 % and q = 30 %. The photographs are scanned
with a pixel size if 15 um which give 7.9 Gbyte of data for
the 36 digital images. With 200 square targets the test field
has been signalized. The size of a target on the ground is
in.
Altogether, 1714 image points of the 200 signalized targets
have to be measured in the 36 photographs of the block.
Most of the signals are imaged in three, six, nine and twelve
images, just one appears in 15 images.
Before we discuss the empirical results of the investigation
we first want to deepen some aspects on the recognition of
simple-shaped signals.
3.1 Aspects on the Recognition of Signalized Points
Theoretical dependences between recognition and shape
of a signal in the context of semi-automatic ground con-
trol point measurement can be solved by simulation. In
fact there exist simulation studies, for example Kaiser et al.
(1992), on the recognition of patterns used for optical posi-
tioning of printed circuit boards. In this work different types
of patterns are evaluated as shown in figure 4.
structures with
repetition
@ E44" +H
basic structures line structures
number of
correlation 1 1
maxima
gradientof | 006 007 | 015 010 010 | 017 0.27
autocorrelation
Figure 4: Features of the autocorrelation function (adopted
from Kaiser et al., 1992).
293
