Analysis of ROV video imagery for krill identification and counting under Antarctic sea ice

An off-the-shelf SeaBotix ROV was deployed under Antarctic sea ice near Palmer Station Antarctica during the September-October 2007 Sea Ice Mass Balance in the Antarctic (SIMBA) project from the research vessel NB Palmer. Video imagery taken showed significant numbers of Antarctic krill (sp.
Euphasia superba
and/or
Euphasia crystallorophis
) under the sea ice at the two stations deployed. The goal of this image analysis is to estimate the krill population densities, as well, as to identify other life forms. The relative motion between the krill and vehicle complicate the video analysis process. The avoidance behavior of the krill adds to this challenge along with the changing lighting conditions under the ice. We discuss these challenges and the algorithms that are under development in this paper. The ROV videos were converted into a string of images, which were used to simulate a running speed of 3 frames/second (fps). A 5 second clip (15 frames) was selected as an initial test for the vision software. The LabVIEW
^reg
Vision Builder AI software has been selected as an image processing platform, which allows us to rapidly prototype algorithms. We have applied noise reduction techniques to reduce some of the noise. Edge detection filters (such as, but not limited to the Roberts Filter) have been applied to further reduce the image's noise level and to increase the contrast between the krill and the water. Next, we applied thresholds to detect the object, which was subsequently used to identify and count and log the number of distinct objects in the image. Once all of the parameters were set, the 15 images were cycled through the configured inspection in chronological order to simulate an actual inspection of the ROV video. The results of our automated krill population estimate to that of humans counting the krill manually using the same video was conducted. We found that illumination and image quality allowed the most prominent individuals in proximity of the camera to be counted. However, because of background noise and low scattering of some individuals the filtering removed suspected individuals that would suggest this krill density is significantly underestimated. The technique however appears practical. In considering the motion of the krill relative to the vehicle, tracking becomes paramount. Plans for implementing this technique in ongoing development are discussed.