Williams et al.: Use of stereo-camera systems in assessing rockfish abundance and pollock behavior 
355 
measurement of targets (Shortis et al., 2000). The 
housings were mounted side by side on the aluminum 
frame (Fig. IB). 
Illumination was provided by two lights mounted 
above the camera housings inside the aluminum frame 
(Fig. IB). The lights were 50-watt high-intensity dis- 
charge (HID) Xenon lights with 5300 lumen output 
and 3900 Kelvin color temperature. The lights were 
inserted into 3-in (7.62-cm) diameter titanium hous- 
ings and the entire light weighed 5 lb (2.27 kg). The 
lights were powered by a battery located in the camera 
housing and linked to the light housing by underwater 
connectors. Four rechargeable 4 Ah 12 V nickel-metal 
hydride batteries were connected in parallel to provide 
approximately 1.5 hour of light per deployment. Each 
light housing was mounted on an adjustable mount that 
allowed even illumination of the target. 
Observing fish behavior in a trawl 
with the still-frame system 
The still-frame system was designed to be light and 
small enough to be easily attached to the inside of a 
survey trawl without significantly changing the fishing 
activity of the net. The system also needed to provide 
adequate illumination and resolution in order to allow 
the fish inside the net to be observed at a range of up 
to 6 m as they passed though a midwater survey trawl 
40 m ahead of the codend. A pair of Canon Rebel Xt 
8 megapixel digital single-lens reflex cameras (Canon 
USA, Lake Success, NY) were used to capture fish 
images. Both cameras were outfitted with 4-gigabyte 
compact flash memory cards for storage of the images. 
A high-power wide-angle Xenon strobe (90°, 150 W/s) 
was used to illuminate the field of view. Three 4-Ah 12 
V batteries were mounted in the strobe housing; two 
were connected to the strobes and the third was used 
to power the cameras. 
The cameras were mounted in separate housings 
made from 10-in (25-cm) diameter deep-water-rated 
(1800 m) trawl floats. Images were taken though a 25- 
mm thick flat acrylic viewport. The strobe and batteries 
were mounted in a third float housing (Fig. 1C). All 
three float housings were secured on a sled constructed 
of 25-mm thick plastic plate and aluminum rails for 
protection. The approximate weight of the complete as- 
sembly in air was 30 kg and was positively buoyant be- 
cause of the float housings. Quick-release trigger snaps 
were attached to the ends of the plastic mounting board 
for attachment to the inside of the trawl. The cameras 
were aimed across the trawl, perpendicular to the wa- 
ter flow to provide lateral views of fish passing by. The 
trigger on the camera shutter was controlled by using a 
microprocessor that was programmed for the study and 
that located in one of the camera housings. A two-axis 
tilt sensor was attached to the microprocessor board to 
allow measurements of fish tilt (deviation of snout-tail 
axis from the horizontal) and yaw (angle of fish head- 
ing in the horizontal plane) to be adjusted from being 
relative to the camera platform to being in absolute 
orientation. A pressure switch was used to activate the 
system once the depth exceeded 20 meters. Images were 
taken at intervals of 5 s to reduce the influence of light 
on fish behavior and to ensure that a new group of fish 
was observed in each frame. The system was capable 
of taking about 400 images or operating for 33 min of 
trawl time per deployment. 
Calibration of the two types of stereo cameras 
The same calibration procedure was used for both stereo- 
camera systems. The basic procedure required collecting 
images of a target plate with a printed 10x10 square 
checkerboard pattern of known dimensions (50x50 cm 
squares for the video-drop system, 100 x 100 cm for the 
still-frame system). This calibration was performed 
underwater. The video-drop system cage was suspended 
in the water while the research vessel was secured to 
the dock. The approximate depth of the camera was 
1 m and the distance from the target was 2 m. The 
checkerboard target was lowered into the water along 
the vessel until it was plainly visible in both cameras. 
The target was then slowly moved horizontally and 
vertically through the field of view of both cameras. 
Up to 15 min of calibration video was collected by this 
method. For the still-frame system, an external trigger 
cable was attached to the assembly, and the system 
slowly moved about while capturing images of the fixed 
checkerboard plate. 
To calibrate the video-drop system, progressive scan 
video images were collected at 29.97 frames/s in each 
camera, and the beginning of the video feed from each 
camera was aligned by using a light-emitting diode 
(LED) synchronization light at the beginning of de- 
ployment. This process was repeated at the end of the 
deployment to confirm that the video frames were still 
aligned. For the calibration procedure, still frame im- 
ages were extracted from the aligned video at 1-s in- 
tervals with Adobe Premier software (Adobe Systems, 
Inc., San Jose, CA). Synchronization was not necessary 
for the still-frame system because the cameras were 
triggered simultaneously. Approximately 20 paired im- 
ages where the target checkerboard was visible in both 
cameras were randomly selected for the calibration of 
each camera system. 
The calibration parameters were estimated with the 
camera calibration toolbox in Matlab, a freely available 
software analysis toolbox built with Matlab computing 
language (Mathworks, Inc.; Bouget, 2008; Fig. 2). For 
each image pair, the position of the corner points of the 
checkerboard pattern were identified by clicking on the 
images and the location of these points in the still im- 
ages was computed by the calibration software to deter- 
mine the intrinsic parameters of each camera. Intrinsic 
parameters were used to correct the individual images 
for optical distortion resulting from the camera lenses. 
The checkerboard pattern allowed the software to auto- 
matically pinpoint exact corner locations based on the 
color contrast of the square boundaries, making the 
initial precision of the manual clicking less critical. 
