The use of aquatic animals for research purposes has been increasingly popular in recent years and this is especially true for the zebrafish (Danio rerio), which is being used for genetic, environmental and pharmaceutical studies, among others. The motion trajectories of the fish can be an important metric and the studies are therefore often supported by a computer vision system, which automatically tracks the fish. However, the tracking is often only conducted on a single fish or in two dimensions, as most of this type of tracking software has been developed for mice, rats or similar. The focus of this thesis is hence to investigate the possibilities of tracking zebrafish in three dimensions.

The proposed system is made of off-the-shelf hardware and consists of a stereo-vision setup with two GoPro HERO5 cameras. The zebrafish are detected in each camera using the SURF keypoint extractor after which the 3D positions are estimated. The 3D reconstruction is based on ray-tracing in combination with Snell's law, in order to account for refraction. By using a Kalman filter and Munkres algorithm it is possible to assemble tracklets with an average length of 150 frames.

A final part of the system, responsible for linking the tracklets, has not been implemented in this iteration of the system.