Zebrafish (Danio rerio) has become increasingly important in medical trials due to the neurobiological proximity between humans and zebrafish. Studies, where zebrafish are exposed to various medications and drugs and their behavior, are analyzed. In such trials, a precise and accurate description of the fish movement in the tank is needed. Tracking of zebrafish has become an active topic of research and is supported by computer vision to accommodate the need for movement trajectories.

In this project, a system for automatic tracking for tracking multiple zebrafish is proposed, based on multi-object tracking state-of-the-art methods within pedestrian tracking. A framework exploiting the regression head of the object detector is used for generating 2D trajectories in two views, top and front, which are used to reconstruct and estimate a 3D position. A large amount of identity swaps is found to be harmful for the 3D triangulation module, return poor performance on the benchmark dataset. A method for detecting and minimizing identity swaps in a given sequence, based on sharing information between views in the sequence.

Results show great potential for a method to detect identity swaps, where the current method for correcting the detected swaps is not supported by the obtained results. More work should be invested in finding an optimal solution for modifying the tracklets around the ID swap detections.