Multi-Object Tracking for Animals with Evolutionary Computation Based Optimization

To analyze animal navigation, tracking of animals is important. The fish tracking challenge is a newly introduced challenge that aims to accurately detect ten fish and track them. For that purpose, the SweetFish dataset was provided. To develop a tracking system of fish, we finetune an object detector on SweetFish dataset to detect fish bounding boxes, because a pre-trained object detector cannot accurately find animals. For tracking, deep-learning based tracker was used and their parameters were automatically optimized by evolutionary computation, because re-identification cannot be used for fish detection and it is necessary to optimize hyperparameters for this case.