Imaging sonar-based fish detection in shallow waters

Traditionally, fish stock assessment is a time-consuming, expensive, and invasive task, since fish are caught and counted using research vessels. Therefore, in a joint project a non-invasive, acoustic-optical Underwater Fish Observatory (UFO) is proposed and developed. The UFO counts and classifies fish utilizing a stereo camera system and a sonar system in order to observe the available biomass. In this paper, we first give an overview of the whole system. Afterwards, we focus on the acoustic part of the system, which utilizes an imaging sonar in order to detect, track, and classify fish. The sonar is mounted on a fixed lander and collects data over long periods of time for the study of fish behavior. It is also intended to trigger cameras for more detailed analysis, when fish are detected in the visibility range of the cameras. The work describes the first processing steps of sonar images for parameter estimation and segmentation. The performance of the fish detector is evaluated using a set of manually annotated sonar images. The detection results show good agreement with the annotations, although some challenges for future improvement of the detector still remain.