4. RESULTS
Experimental results show, that the models discussed above
have significant differences. Different combinations of fixed
parameters and different combinations of integrated
observations occur to be distinguished. In the following sections
the differences are shown. Beginning with the most general case
including as many observations and as many parameters to be
estimated as possible, the number of estimated parameters and
observations are stepwise reduced.
4.1 Reproducibility
To evaluate the usefulness of a certain model, the concept of
reproducibility of the estimated parameters is used. The
reproducibility is measured by the standard deviation of the
results of a certain model using distinct sets of observations.
While evaluating the reproducibility it has to be taken into
account that the calibration parameters may be disturbed by an
interruption of a single session, if an user takes off the glasses
and puts them on again. In a session, the relative position and
orientation of the user's head (measured by the tracker) and the
glasses should not change. For this reason, the reproducibility
can only be checked by a random sampling of a larger set of
observations of the same session.
4.) Number of estimated parameters
First of all, it is searched the best constellation of the estimated
parameters. All available observations are used and their
residuals are estimated. The unknown parameters describe the
transformation from the sensor co-ordinate system to the
display co-ordinate system. Where:
C; ,C7- focal length in row and column direction of the
image co-ordinate system in pixel,
Xo, o = principal point in pixel,
X0 Yo Zo = components of the translation from the
sensor system to the eye system in meter,
O, ¢, K = components of the translation from the
sensor system to the eye system in radiant.
Par. Results
absolute standard deviation standard deviation
value (A) (A) (B)
€; 1218.05 159.68 32.4118
C2 1054.42 140.58 6.8377
Xo 215.09 34.33 0
Vo 236.91 37.40 0
Xp 0.0347 0.0130 0.0068
Yo 0.0325 0.0257 0.0035
Zp 0.0537 0.0571 0.0085
e 0.7782 0.3574 0.0764
9 1.6918 0.0298 0.0077
K 1.0981 0.2929 0.0813
Table 1. (A): all observations used and 9 parameters estimated.
(B): all observations used and constant principal point.
In table 1 an example for a calibration result is given for two
cases in which all observations are used. The reproducibility is
rather weak for case (A). The co-relations between the
parameters reaches up to 99%. That means, there are too many
unknown parameters introduced into the model. Several
experiments showed that the best strategy here is to keep the
principle point fixed at the centre of the image. It can be
observed from the values of the fourth column of table 1, that
there is a significant increase in reproducibility.
4.3 Number of observations
In contrast to the section 4.2 the number of observations is
decreased in the following. Firstly, the second bird is left away,
secondly, the first bird is assumed to be error free. Then, using
both birds, the number of image points is reduced consequently.
In table 2 can be seen that there are significant differences
between the conventional methods and the method proposed in
this paper. Comparing column four in table 2 with column four
in table 1 it can be observed that there is a decrease of accuracy
when assuming the tracker to be error free (B). This is the
assumption made implicitly when only image points are
assumed to be affected by errors. The influence of the tracker
errors can also be seen in column three of table 2. When
assuming only the pixel co-ordinates to be error free and only
the tracker errors are considered (A) the results are still better
than (B).
The last table 3 gives an impression how much points necessary
to get reliable values. In the case of using only ten points, the
focal length has to be assumed to be constant. Otherwise, there
will occur cases that have no convergent solution when using
the Newton method. The table 3 shows that the accuracy of the
focal length decreases faster than the accuracy of the parameters
of the transformation form the eye-system to the display-
system.
Par. Results
absolute standard standard
value deviation (A) deviation (B)
€; 1388.9982 71.8874 98.2841
C2 1359.0145 54.5555 77.1293
Xo 215.09 0 0
Yo 236.91 0 0
Xo 0.0364 0.0074 0.0053
Yo 0.0487 0.0077 0.0107
Zo 0.0449 0.0443 0.0482
1.9015 0.1411 0.0679
Q 1.6715 0.0150 0.0221
K 1.7994 0.1462 0.0632
Table 2. (A): second bird left away. (B): the tracker is assumed
to be error free.
Param. standard deviations
obs. num. 10 20 30
cy 0 103.1610 32.4118
€? 0 79.1103 6.8377
XQ 0 0 0
yo 0 0 0
X» 0.0059 0.0046 0.0068
Yo 0.0110 0.0093 0.0035
Zo 0.0127 0.0495 0.0085
œ 0.1127 0.0787 0.0764
0 0.0142 0.0098 0.0077
K 0.1160 0.0765 0.0813
Table 3. Examples using different numbers of observed image
points. A zero standard deviation indicates a fixed parameter.
—538—
[4