Computational techniques for automated tracking and analysis of fish movement in controlled aquatic environments

This thesis presents research on automated video analysis using computer vision systems, for analysing fish movements and behaviours in sea cages and tanks. Video technology is widely used in aquaculture to observe fish movements, however these observations can be subjective and fish can only be observed for short periods due to the manual labour required and observer fatigue. In research, the analysis of video footage is tedious and very time consuming, requiring sampling to make it feasible. It is therefore desirable to automate such analysis and provide users with a tool that can gather data about fish movements in real-time, continuously, objectively and at a high sampling rate. The aim of this thesis is to develop and validate computer vision systems to track fish automatically in sea cages and tanks. Three computer vision systems are proposed, one for sea cages and two for tanks, and they consist of two major stages. The first stage extracts fish images from complex backgrounds in video footage through the process of segmentation. The second stage is responsible for tracking multiple detected fish by associating newly extracted objects with existing tracks of fish. The system developed for use in sea cages tracks fish for short periods and generates measures of fish movement - their average swimming speed and direction. The first system used in tanks tracks a small number of fish over a long period of time with the purpose of longterm observation of spatial location and agonistic behaviours between individuals. The second system used in tanks is based on the one developed for sea cages and is used to track small groups of fish, with the purpose of observing groups' spatiotemporal patterns rather than movements of individuals. When using the sea cage system, variations in swimming speed and direction were observed within days and between days. Some of these variations could be attributed to water current changes due to tides, but no consistent patterns were observed in relation to time of day or feeding. During the transfer of Atlantic salmon smolts from a freshwater hatchery to sea cages, a pattern of non-schooling behaviour during the first 3-5 weeks was observed, followed by a sharp transition to schooling behaviour. In tanks, tracking of two individuals was possible but maintaining unique identification of fish was not completely achieved. When tracking groups of fish, the tracking system was able to detect variations in swimming speed, while spatiotemporal patterns were observed in relation to the demand feeder and the water inlet. The sea cage system has a potential application in the commercial setting, where it can be used to develop behavioural profiles of fish and act as an alarm system if unusual behaviours are detected. From the research point of view, use of these automated systems improves the process of gathering data about fish movements, provides a high level of sampling and increases the speed of video processing, which is currently based on manual observation of fish movement. Time can be spent on analysing the data rather than on extracting it from video. While there is some requirement for data analysis in this system, the benefits in extracting data from the video automatically far outweigh the requirement for data analysis.