4-1-6 
• the interior orientation and distortion parameters for each 
camera, 
• the relative orientation between the cameras, and 
• the EO of the dual-camera unit with respect to the 
GPS/inertial sensors which define aircraft position and 
orientation. 
It is well known that in order to recover camera interior and 
exterior orientation via self-calibration, in a situation where the 
object point field is not well distributed in three dimensions, two 
fundamental requirements must be met. The first is provision of a 
convergent imaging configuration, whereas the second involves 
inclusion of orthogonal camera roll angles in the photogram- 
metric network design. With the airborne RCAMS, convergent 
imaging geometry can be attained quite simply by flying over a 
target array from different directions since the Pulnix CCD 
cameras are both mildly convergent with respect to one another 
and have a downward pointing angle of 35° from the horizontal. 
At the low-level operating flying height of 50m the ground 
coverage of each image was approximately 40m x 35m. A 
practical self-calibration scenario could therefore be formulated 
using a calibration range of only about 50m x 50m on the ground. 
By flying over this range from the four cardinal directions, as 
well as from the NW, NE, SE and SW, the convergent imaging 
geometry illustrated in Figure 3 could be realized. Given that the 
cameras are mounted on the wings, the prospect of providing a 
diversity of camera roll angles presented some practical 
difficulties. These were overcome by first mounting the cameras 
upside down and recording a set of images, after which the 
cameras were re-mounted in their correct, upright orientation. 
The final calibrated EO related to the second camera 
configuration. 
Although the provision of a 180° kappa-angle diversity did 
facilitate an adequate recovery of camera interior orientation 
parameters, the absence of a 90° roll angle hindered the recovery 
of any differential scaling (affinity) introduced into the imagery 
by the video image acquisition system. Nevertheless, initial 
analysis of the calibration data suggested that a sufficiently robust 
recovery of camera system parameters and EO could be achieved 
with the adopted imaging arrangement. 
A considerable number of individual synchronized image pairs 
were available for the multi-sensor self-calibration of the airborne 
RCAMS stereo imaging system as a consequence of the recording 
of video sequences. The final photogrammetric network adopted 
comprised 12 image pairs, eight as indicated in Figure 3 (about 
half of which had a 180° roll angle), and four from the cardinal 
directions at the higher flying height of 70m. Thus, the final self 
calibrating bundle adjustment comprised two cameras and 24 
images. 
3.2 Control Point Array 
The object array itself comprised a grid of 15 x 15 targets at 
approximately 3.5m spacing. Under normal circumstances 
governing camera self-calibration with strong network geometry, 
ground control information would not be required if the aim were 
restricted to dual-camera self-calibration and relative orientation. 
In this instance, however, it was imperative that camera EO be 
determined in the same coordinate reference system as the 
GPS/inertial system used to position the aircraft. All ground 
points were surveyed with a total station to about 30mm relative 
precision and to about 0.1m absolute accuracy via GPS surveys of 
a number of the points. 
3.3 Image Recording and Mensuration 
Under the normal operating scenario for airborne RCAMS, time 
tags corresponding to an aircraft interval of 20m, or about every 
15 th frame, are written to the video tape. In subsequent extraction 
of the desired image pairs from the SVHS video tape, these 
tagged images are converted to digital form, compressed via 
JPEG with a compression ratio of approximately 17:1 (i.e. to a 
file size of around 20 kb) and written to CD-ROM as geo- 
referenced images. 
Within the self-calibration, account had to be taken of the 
possible perturbations to the original video images arising from 
this analog-to-digital conversion, which also involved an 
effective resampling of the imagery to a resolution 736H x 560V. 
To allow for any differential scaling within the analog-to-digital 
conversion, an affinity term was carried in the self-calibration 
model (Eq. 1). Also, the principal distance derived within the 
bundle adsjustment would compensate for any common scale 
error in assumed pixel array dimensions arising from the image 
conversion process. 
An important component of the airborne RCAMS is the DVS (for 
Digital Video System) which allows an operator to access an 
image pair from CD-ROM by time, geographic location, or 
powerline attributes (e.g. pole number). Once the stereo images 
are displayed side by side on the PC monitor, the spatial position 
of any point of interest in the scene, be it a power pole or 
vegetation, can be determined through simple intersection once 
the interior and exterior orientation of the cameras is known. 
Figure 4 illustrates the display presented to the operator of the 
DVS. A zooming facility is incorporated for more precise 
pointing to image points, as was warranted in the measurement of 
the target points of the calibration range. 
Although the camera calibration range comprised 225 targets, the 
number measured in each of the 24 images averaged 90, with the 
range being from 40 to 170. Only targets that could be reliably 
observed to a precision of 1 pixel or better were recorded. As an 
immediate check on the validity of the image measurement 
process, spatial resections were carried out using all available 
control points and the nominal calibration data for the two Pulnix 
CCD cameras. In spite of possible adverse influences on camera 
calibration parameters, such as the fact that each camera was 
housed behind a glass viewing port, all resections yielded RMS 
image coordinate residuals of between 5.8 pm and 8.6 pm, or just 
under 1 pixel. This was an encouraging sign for it offered early 
verification that the data was free of gross errors and that the 
desired relative positioning accuracy of the stereo video system of 
0.15m, corresponding to 1 pixel precision, could be attained.