To get information about an underwater scene, the first and most important step is the automatic recognition of living creatures and important objects. Based on this information, the objects can then be measures or tracked with other algorithms in order to extract higher-quality information. A sophisticated background subtraction method has been developed for the segmentation of underwater scenes. In our software, this can be combined with one of several self-developed spatial models. These spatial models ensure that the smoothness of natural images can also be reflected in the segmentation and thus the contours of the objects can be better imaged.
The analysis of behavior additionally requires the assignment of segmentations from different images to a special object, so as to get temporal data on the behavior of this object. In the demonstrator, this assignment is made only on the basis of the segmentations from the previous steps. The method developed for this is independent of the object type and does not require any learning phase.