Insect Tracking in Highly Crowded and Dynamic Environments

Ants are a model system used for social behaviour experiments. Tracking individuals is a critical step when conducting experiments on the social behaviours of ants. Existing ant tracking methods are either labour-intensive or struggle to achieve high accuracy in crowded and dynamic settings. This thesis introduces deep learning-based approaches to enhance tracking accuracy in complex environments. It suggests joint-detection-and-tracking as a solution for ant tracking, offering the best tradeoff between accuracy and efficiency. Additionally, a semi-supervised tracking framework is presented to address changing experimental conditions. These methods are assessed using datasets and benchmarks proposed in the thesis, and real-world experiments.