56
which in theory completely excludes problems such as line jitter and transmission errors. At the time of the experiment a
monochrome version of this camera was not available and a colour version was tested, however it was proposed that the
GRGBGRGB.... horizontal line sequence on the sensor might be deconstructed from the 24 bit RGB image produced by
the camera. For the sake of expediency, the 24 bit colour images were simply band-averaged to 8 bit grey scale images.
Unfortunately this double re-processing of the raw pixel values proved to be not a true test of the sensor.
DATA ACQUISITION AND PROCESSING
Eight frames of a targeted test field were taken with each camera. The geometry of the network adhered to the optimal
ring configuration and 90° convergence proposed by Fraser (1986), in every case. The camera was rolled about the
optical axis by 45° after each frame. The camera to object distance was set in each case as a compromise between
maximum coverage within the format of the sensor and minimal target image losses at the edge of the frame due to
changes in perspective and camera roll.
The target field is a black aluminium plate with an array of 121 passive targets, which are a mixture of targets flush with
the surface and stand-off targets on rods. The target field encloses a volume of 0.5m by 0.5m by 0.25m. Passive
targets were adopted to avoid the response fall off of retro-reflective targets at high incidence angles. In every case the
cameras and the test field were allowed at least two hours to reach thermal equilibrium before the calibration test
commenced. Data acquisition required, on average, 20 minutes. Typically the laboratory temperature is stable to better
than 1°C over such short periods of time, hence it is extremely unlikely that the plate would change significantly in shape
during the time required for a calibration test.
Using the same focus setting and camera to object distance, three or four frames of a plumb line calibration range were
taken with each camera. Again, the camera was rolled about the optical axis by 45° after each frame. The plumb lines
are white plastic cords, under tension, against a black background. A mixture of overhead and auxiliary lighting was
used for the targeted object and the plumb lines. The grey levels were monitored to maximise the target or plumb line
signal, and minimise the background, whilst avoiding saturation of the sensor. Pixel intensities in the target field frames
averaged 200 and 5 for the target images and the background, respectively. The target images spanned 7 and 30 pixels
on average for the low and high resolution sensors respectively. Pixel intensities in the plumb line range frames
averaged 150 and 20 for the plumb line images and the background, respectively. The plumb lines spanned
approximately 4 and 12 pixels on the low and high resolution sensors respectively. Typical images are shown in Figure 1.
Figure 1. Target and plumb line images from (left to right) low, medium and high resolution sensors.
All image windows are from near the centre of frame and are 40 pixels square.
The target images are located within the frame using a semi-automated procedure based on the known object space
coordinates of the targets and a resection/intersection computation (Shortis et al, 1991). Each 8 frame data set
comprised approximately 900 target images. Positions on the plumb lines were gathered using a semi-automated
procedure based on simple line following. Measurement of each frame produced approximately 20 locations on 20-35
plumb lines, leading to a total ranging between 1500 and 2500 image locations from the four frame sets.
Because of the high signal to background ratio and the uniformity of the background, thresholds were set by a simple
additive constant of 5 grey levels above the background. The weighted centroid of the target or plumb line image was
then computed using the non-zero intensity pixels inside the window (Trinder, 1989). More complex image location
schemes, such as template matching or intensity profiles, were not warranted here because of the binary nature of the
images supplied by the targets and plumb lines (Shortis et al, 1994).
Initially, only one set of frames was acquired with the Kodak Megaplus 4.2 camera, using the 24mm lens. During the
processing of the images it became apparent that there was read-out noise in the frames of the test field. The "comet
tails" were not readily visible in the original images and were only shown clearly in a contrast enhanced image. A number
of strategies were employed to overcome the affect of the noise, such as different threshold algorithms and ellipse
fitting instead of centroiding, but no improvement could be achieved in the subsequent calibration computations. A few
months later the camera was again loaned to LaRC and another set of images was acquired, in this case with a 50mm
lens and only of the target field. No significant read-out noise could be detected in these frames.
The image data from the CCD cameras was processed using the CRAMPA suite of programs (Shortis, 1989) executed on
an IBM compatible PC. All plumb line calibration data sets were initially processed using a principal point at the centre of
the frame and no high order additional parameters to model non-linearities or sensor unflatness. All target field data sets
were then treated as self-calibrating free networks, each with an implicit datum supplied by internal constraints, and
explicit constraints supplied by the lens distortion parameters from the plumb line data. In no case were there
significant high order additional parameters, most probably due to the relatively narrow fields of view for all cameras.
The new principal point position was then adopted for the re-computation of the plumb line data set. The target field
network was then re-computed as before. In one or two cases the iterative process had to be continued, either because
the principal point location was significantly shifted from the centre of frame, or because of oscillation of the
parameters. The convergence criterion used was no significant changes to calibration parameters other than the lens
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995
IA