The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
739 
An independent check on accuracy is provided by the length 
measurements on the rigid arms of the frame. The RMS error 
of the distances indicates the base line integrity of length 
measurements made with the stereo-video system. Based on 
centroid measurements, the RMS value is typically 0.05 mm, 
signifying a high level of accuracy demonstrated by the self 
calibration measurements. 
The relative orientation of the cameras is derived from post 
processing of the locations and orientations of all synchronised 
stereo-pairs in the photogrammetric network. Rigid mounting 
of the camera housings to the frame and a rigid connection 
between the cameras and the view ports generally ensures the 
stability of the relative orientation of the cameras (Shortis et al., 
2000). Experience has demonstrated that a weakness of the 
implicit model for refraction is the integrity of the full light path 
from the first water-port interface through to the image sensor. 
A consistent spatial relationship between the view port and the 
camera lens is critical to this stability. 
For subsequent measurements in the field, the photogrammetric 
network provides the required calibrations and the relative 
orientation of the stereo-cameras. The system is then validated 
in the pool environment by introducing a known length which is 
measured manually at a variety of distances and orientations 
within the field of view and expected working range of the 
system (see figure 3). The RMS error of these validation 
measurements is typically less than 1 mm over a length of 1 m, 
equivalent to a length accuracy of 0.1%. This is a best case 
scenario in conditions of excellent water clarity and high 
contrast targets. Experience with shallow water measurement 
of fish silhouettes in more realistic conditions, together with 
validated measurements of live fish in the field, indicate that 
length measurements will have a field accuracy of 0.2% to 
0.7% (Harvey et al., 2002, 2003, 2004). 
3.2 Deep Water Operations 
For deep-water operations there may be measurement 
inaccuracies resulting from the application of a camera 
calibration carried out in shallow water to imagery gathered at 
much greater depths. Stereo-camera calibrations are generally 
carried out at depths of 1-3 m for operational convenience, 
however the stereo-cameras can subsequently be deployed to 
depths of up to 2,000 m. Under these conditions of 
considerably increased water pressure and decreased 
temperature it is expected the camera housings and view ports 
will deform, and the deformation may adversely affect camera 
calibration and subsequent stereo measurement. 
Initial testing for the effects of depth have clearly indicated that 
there is an impact on the calibration of the stereo-camera 
system. The first experiment used continuous calibration based 
on a laser array system (Shortis et al., 2007). Measurements to 
a depth of 500 m has confirmed the presence of significant 
systematic errors in the calibration, however the test did not 
include an independent scale determination. A second 
experiment was based on a scale bar attached to the towed body 
so that it appeared in the edge of the field of view of the 
cameras. A range of distances on the scale bar were measured 
at every 100 m of depth whilst the system descended to 1120 m 
and returned to the surface over a period of 110 minutes. 
Variations of up to 8 mm over a length of 1.2 m, corresponding 
to an error of 0.8%, were recorded. Current research is 
analysing the effects of pressure and temperature on the camera 
housing so that these effects will be fully understood and 
appropriate modifications to the housings can be implemented. 
4. MEASUREMENTS FROM VIDEO SEQUENCES 
Stereo-video images enable accurate 3D measurements of point 
locations. Distances, areas and volumes can be derived from 
these measurements and used to characterise marine fauna 
(figure 4) and seafloor habitat features such as boulders, 
crevices and ledges. These fine spatial scale metrics 
complement information typically gathered at coarser scales by 
techniques such as acoustic mapping (see figures 1 and 7). 
Similar stereo-video techniques were originally developed for 
measuring the lengths of fish to estimate population size 
structure (Harvey and Shortis, 1996) and are based on operator- 
identified points of interest in the stereo-images. 
Figure 4. Example of an operator measurement of the height of 
a deep-water coral. 
Because manual measurement and analysis of large volumes of 
video sequences is time consuming, labour intensive and 
therefore costly, there is considerable potential benefit in 
automating measurement processes. For example, CSIRO 
researchers collect 100+ hours of video recordings annually 
during biodiversity and fishery habitat surveys. Currently, the 
automation techniques employ motion analysis, image 
segmentation against the background, and colour matching to 
identify the presence and percentage cover of benthic fauna, 
and differentiate habitat types in video sequences (figure 5). 
Initial results show promise for rapidly quantifying the cover of 
complex structures such as the reefs formed by stony corals (see 
figure 8), but tuning and validation against manual 
identification techniques remains a work in progress (Williams 
et al., 2008). Stereo-measurement can then provide the sizes of 
individual animals or substratum features within selected image 
pairs to estimate population characteristics. 
Figure 5. Candidate region of stony coral detected within an 
interest window (grey trapezoid) from motion analysis.