Juergen Oberst
A first attempt was made to measure the stars' positions manually. This was accomplished using an appropriate tiepoint
program which allowed us to make the line/sample measurements at sub-pixel level. Stars seen in the in images were
correlated with known stars by comparison with charts and catalogs.
Next, the star identifications and position measurements were made automatically. For this purpose, an image area mi
having the size, for example, of i=20 x j=20 pixels around some "pilot" star was extracted and placed into a mask.
Subsequently, the entire image was scanned for windows wij, to match this particular pattern. A star identification was
considered to be acceptable, when the cross correlation coefficient c, given by the expression c = cp / Ve / Ver,
exceeded a fixed value, e.g. 0.5, for a given window wjj of the image. cp, c], and cp were then computed by
multiplication of corresponding DN values of the windows mij and wjj (from which the mean DN values for the
respective windows was subtracted) and by summations over all lines and samples, i,j:
co = à wij * mij (4a)
€1 = mij * mij (4b)
c2 = 2 Wij * wij (4c)
The correlation coefficients c for each window position were stored in data arrays, in which the position of the star was
determined by searching for local maxima. The coordinates of this maximum were determined with subpixel accuracy.
From the right ascension and declination of the stars in the catalog, we computed the expected line/sample coordinates
of these stars using the nominal pointing data available for each image. These were compared with the star positions
from the automated star search. Normally, 50-80% of the stars in the chart, falling within the area of the respective
image, were successfully correlated with stars in the images.
In the following, the precise positions of the stars were compared with the catalog. A least-squares program was
developed to determine the camera focal length and the precise pointing of the image in terms of boresight right
ascension, declination, and twist on the basis of the identified stars. The nominal image pointing parameters were used
as a first guess to start the iterations. The 1- o value of the residuals of the star positions after these fits are shown in
Table 4. While 6 — 1.07 pixels, on average, for the manual fits (second column), this value is effectively reduced to
G =0.75 for the automated matching (third column). The large o (^1) for some of the manual measurements is
probably due to misidentifications of stars. When the identification of stars is correct, the accuracy of the
measurements is typically on the same order as that of the automated matching. When distinct signatures of smeared
stars are available, the automated matching is clearly superior to the manual measurements (see image starcl131, Fig.
4). In cases of very heavy smear, however, in which star signatures overlap, the automated matching failed.
5.2 Distortion Model
The residuals were then examined for image distortion (see Fig. 6, left) and subjected to the laboratory distortion
function (eqs. 1-3). The residuals are reduced from 6 = 0.79 to 6 = 0.617 (39.1%), a quite moderate improvement (Fig.
6, right). Alternatively, the laboratory function (Table 3) was used to correct the raw line/sample star positions using
equations (1-3); then the fitting was repeated. The residuals after the fit are reduced to an average of 6 = 0.47 pixels,
with some images showing residuals of as small as 6 = 0.32 (Table 4, rightmost column). This indicates that the
characteristic distortion pattern found during the calibration on the ground can still be identified in the in-flight images;
but obviously; the random noise in the star position measurements is large for some of the images. The camera focal
length was found to be: f = 685.24 +/- 0.29 mm (mean of all fits), a value which has a higher uncertainy and which is
slightly smaller than the laboratory value of 686.55 +/- 0.05. Both focal lengths are clearly above the nominal value of
677 mm.
In the next step, an attempt was made to determine a new distortion model from the in-flight data to identify possible
changes in the camera geometric properties after spacecraft launch. All residuals after the individual fits were
combined and new polynomial coefficients was fitted, as was done for the laboratory images. The initial residuals with
G — 0.791 were reduced to 6 = 0.525 after the fit. Thus, the laboratory model accounts for 56% of the scatter in the
data. Although this new model seems to be a better fit than the laboratory model, tests suggest that the two sets of
coefficients probably differ only because of the random errors in the data. Random errors in the line/sample position
measurements are clearly larger than in the calibration images (6 = 0.354).
226 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000.
Ta
Im:
Fil
sta:
sta:
sta
stai
sta
sta
sta
sta
stai
nav
nav
nav
nav
nav
nav
nav
nav
nav
nav
*%)