182
= =
Figure 4: Random pattern with macro structure Figure 5 : Camera setup for calibration
5.EXPERIMENTAL INVESTIGATIONS
5.1 Calibration
As an example the results of the geometric calibration of two Kodak DCS 200 cameras are given. The cameras were
equipped with a 24 mm lense.
The calibration was applied with fixed focus at an object distance of approx. 1.30 m. The viewing area of one camera
is about 0.8 m? at that distance. A fixed pixel size of 0.01 mm was introduced. Figure 5 shows the geometric setup of
one set of images.
Since the parameters a; and a; were highly correlated for this setup, the parameter a; was cancelled in the final bundle
adjustment. Also, the parameters a; and a4 were not significant, i.e. the pixels are quadratic and rectangular. The co-
ordinates of the principle point and the focal length were correlated by 30 % to 50 %. This is mainly caused by the
planimetric point field and the small viewing angle of 37 degrees. Table 1 shows that the calibration becomes more
stable the more points are used for the bundle adjustment. However, measuring a considerable large number of points
tends to be no problem, since the measuring speed is about 8 points per second. Altogether, a calibration of a Kodak
DCS 200 could be completed within one day including preparation, image capture, interactive and automatic point
measurement, and bundle adjustment.
The resulting calibration parameters (c.f. table 2) lead to corrected image coordinates with an accuracy better than
1 um, which is sufficient for the matching accuracy of 1/3 pixel (3 um) for point cloud measurements and 1/10 pixel
(1 um) for profile measurements.
5.2 Controlled tests >
5.2.1 Application with still video camera
In order to confirm the accuracy potential of the calibrated Kodak DCS 200 cameras, a 20 % concept model was digit-
ized with the two measurement options. We set up a stereo camera configuration with a camera-to-object distance of
1.3 m, a base-to-height ratio of 1:2, and a convergence angle of about 20 degrees. Hence, the pixel size on the object
was approximately 0.54 mm. Control points for the orientation of the image pair were provided on the object in ad-
vance by an independent bundle adjustment using several images and distinct reference points which defined the car
coordinate system.
In order to get reliable reference data for the accuracy tests, the object was also captured with a standard analog cam-
era Rollei R6600 equipped with a 40 mm lense. This well-known analog camera was used in a stereo configuration
with a camera-to-object distance of 0.80 m, a base-to-height ratio of 1:1.5, and a convergence angle of about 20 de-
grees. The analog photos were digitized on a PS1 scanner with a pixel size of 7.5 um. Hence, the pixel size on the
concept model amounted to 0.15 mm for that stereo pair. Reference data were finally provided by profile measure-
ments which were derived from the digitized R6600 image pair. Because of the small pixel size of 7.5 um, the measur-
ing precision of a single profile point theoretically amounts ápproximately to 0.04 mm. This seemed to be sufficient for
the subsequent accuracy evaluation of the two measurement options derived from the DCS 200 image pair, because
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995