No abstract

With the development of digital imaging techniques over the last decade, there are now new opportunities to study complex behavioural patterns in fish (e.g. schooling behaviour) and to track a very large number of individuals. These new technologies and methods provide valuable information to fundamental and applied science disciplines such as ethology, animal sociology, animal psychology, veterinary sciences, animal welfare sciences, statistical physics, pharmacology as well as neuro‐ and ecotoxicology. This paper presents a review of fish video multitracking techniques. It describes the possibilities of tracking individuals and groups at different scales, but also outlines the advantages and limitations of the detection methods. The problem of occlusions, during which errors of individual identifications are very frequent, is underlined. This paper summarizes different approaches to improving the quality of individual identification, notably by the development of three‐dimensional tracking, image analysis and probabilistic applications. Finally, implications for fish research and future directions are presented.

The capability of a new multitracking system to track a large number of unmarked fish (up to 100) is evaluated. This system extrapolates a trajectory from each individual and analyzes recorded sequences that are several minutes long. This system is very efficient in statistical individual tracking, where the individual's identity is important for a short period of time in comparison with the duration of the track. Individual identification is typically greater than 99%. Identification is largely efficient (more than 99%) when the fish images do not cross the image of a neighbor fish. When the images of two fish merge (occlusion), we consider that the spot on the screen has a double identity. Consequently, there are no identification errors during occlusions, even though the measurement of the positions of each individual is imprecise. When the images of these two merged fish separate (separation), individual identification errors are more frequent, but their effect is very low in statistical individual tracking. On the other hand, in complete individual tracking, where individual fish identity is important for the entire trajectory, each identification error invalidates the results. In such cases, the experimenter must observe whether the program assigns the correct identification, and, when an error is made, must edit the results. This work is not too costly in time because it is limited to the separation events, accounting for fewer than 0.1% of individual identifications. Consequently, in both statistical and rigorous individual tracking, this system allows the experimenter to gain time by measuring the individual position automatically. It can also analyze the structural and dynamic properties of an animal group with a very large sample, with precision and sampling that are impossible to obtain with manual measures.

No abstract

No abstract