0bservation of digital images on split screen ste-
reoscope system. As far as the acceleration of
measurements is concerned the grafical measuring
mark in the form of a window. The sizes of window
are changed depending on target dimentions of the
image. To make measurements one only needs to
locate the target into corresponding windows in the
right and left images. In this case the measurements
are made faster and it is not necessery for observer
to have much qualification. If the measured point is
nota target the sizes of the mesuring mark should be
about 20x20-30x30 pixels of image, because within
these limits the obtained precision of matching is
à ptimal, according to D.Resenholm,1987.
The following processing of images is caried aut
within the limits of the windo ws.
2. Edge detection is perfomsd here by means of the
convolution of the image with a well known Sobel
operator. As a result we get the gradient image.
3. The circular target location is based on solution
of the equation:
(X; -KY+(Y- L)-K=0, (1)
This equation is formed for each pixel i (within the
window) the gradient of which is not zero.
The robust technique is applied here to compute the
unknown ceurdenates af senlur X.Y and racits H.
in order to reduce the influence of random noise on
the precision of pointing and to narrow the target
edges (till zl pixel) we use the following weight of
eq. (1):
Gm
& - exp(]eze -1b. (2)
where Gmax is maximum value of gradient in the
window; G; is the gradient value of the pixel i in the
window.
To avoid detecting noise (shade, patch of light and
so on) as target edges we use robust estimation
technique, i.e. ihe following weigh! functions are
introduced:
1 it |vis2u
P; =4exp[-0.1C|V;[iu}1 it |V/>2u and N<3 (3)
expl-0.1(|V;|4u)3) it |V;|>2u and N>3
Here V; is the descrepance of i equation (1): pis the
standard error of unit weight; N is the number of
iteration. The computation is iterated until the
required precision is obtained.
4.The cross target location is based on the de-
termination of the target coordinates as a point of
intersection of two lines.
The coefficients of line equations are determined
using the same principle of a robust estimation
technique. Moreover, the pixels are divided inte
two groups (for each line) applying the directions of
its gradients.
The same algorithm is used for determination of
coordinates of a contour point which can bs repre-
sented as a point of intersection of two lines.
5. The square target location is based on the
computation of the square center as the point of
intersection of diagonals, The needed conditions for
the pixels detection which belong to 4 sides of a
square are easily obtained from analicis of the
gradients directions of pixels. Than the same ap-
proach is used for the robust estimation of a line
equation.
These algorithms of tageis location in detail can be
found in A.Chibunichev {(1991,1992).
6. Image matching. First of all it should be said that
the algorithms of target location (above-mantioned)
solve the pointing problem without image matching.
However (as will be shovn below) the matching
process permits to improve the precision of de-
termination of paralaxes for cross and square targets
as well as for counter points.
The image matching can be done with many methods
(M.J.P.M.Lemmens,1988). The least squares mat-
ching method (A.Grun, 1985, D.Rosenholm, 1987,
Heipke C, 1991) was chosen here because it gives
high accuracy potential, high degree of invariance
against geometrycal image distortions and relatively
simple possibilities for statistical analysis of the
results. The disadvantage of this method is a quite
high time-consuming. Some recomendations to re:
duce the computational cost of least square matching
are mantioned in A.Chibunichev, 1992.
Investigation of precision of target location
First, let consider the results of extensive studies of
the precision of circular target location on digital
image f AC hibunichev, T. Shimahaneova, 199735. They
in cgiligatiuno have been Carried uulun ihe basis of
the artificially generated targets with varying char:
acteristics, The simulated process of digitizing was
performed for the following pixel sizes 2.3, 6.8, 7.5,
9, 12.5, 17, 19, 23, 27m, wich correspond to values
for real CCD cameras (T.Luhman, 1990). For each
pixel size eight different quantization levels (grey
scale values) were investigated: 2'.2*.....2 , which
means encoding into 1,2,...8 bits respectively. The
target location was carried out 50 times for sach
pixel size, quantization level and target size (100
and 200 um). Moreover, the exact position of the
target center was changed by a random number up to
+1 pixel for each new digitizing process. The
variation in precisions of target pointing (m,)
depending on the quantization level are illustrated in
figure Q for a target zise of 100,m (diameter of
circle). The value m,- /mè+ my, where m,, my,are
standard errors of target! center coordinates de-
termination. The pixelsizes are indicated on the left
sides of the curves and on the rightsjdes - the ratios
of target size! pixel size. Other target size/ pixel size
ratios were studied, but the precisions of taget
location in these cases were approximately similar to
those shown in figures 2 and 3.
Figure 3 illustrates the influence of the image random
noise on the precision of pointing. It should be
painted out, that the random noise was introduced
into the values of the pixels (during the simulated
process of digitizing), but prior to quantization .
The random noise percentages were *5*€, 210%,
+15%, £205, z40% which equivalent to ratios
K»signal/soise of 20:1, 10:1, 6.7:1, 5:1, 2.5:1. The
figure 3 corresponds to the case when a quantization
level is equal to 256 (2%).
The figures 2 and 3 demonstrate that a better
precision of target location (near 0.01 pixel size)
can be obtained when the quantization level is equal
io or greater than 32 (2%), target sizes are larger than
6*pixel size and the noise is less than 10% (K=1:10).
The similar results were o bteined in J.C.Trinder,1989.
