Automated Change Detection in an Undersea Environment using a Statistical Background Model

Marine scientists are turning increasingly to underwater video cameras in their research. These provide enormous quantities of visual data that often overwhelm the manual processing abilities of the scientists. To cope with such large data sets, an automated change detection system is proposed that helps isolate the time periods in which significant activity is found in the video sequence. Unlike change detection algorithms in use in terrestrial environments, the system must account for the photometric complexity of underwater video, including interference from small floating particles ("sea snow"), the scatter of light as it propagates through water, and the non-uniform frequency decay of light intensity with distance. In addition, certain activity, such as the motion of swimming fish that are attracted by the use of artificial lighting, is considered a distracter, and should, ideally, be ignored. These factors are addressed by our system, in large part through the use of Mixture-of-Gaussians background models.