Zebrafish Object Tracking: A Comparative Evaluation of DeepSORT and ByteTrack

Sammanfattning

Zebrafish are widely used in laboratory research, but manual monitoring remains laborintensive and prone to error. This thesis evaluates two different multi-object trackers, Byte- Track and DeepSORT, on their abilities to accurately track zebrafish. Integrated into a Python pipeline, the system was evaluated on a variety of aquarium video clips under different resolutions, frame rates, and lighting conditions, as well as on an annotated ground-truth sample. Results show that motion-based tracking (ByteTrack) offers higher throughput and precision, while appearance-based association (DeepSORT) yields slightly longer continuous tracks. However, common challenges - reflections, occlusions, and erratic swimming behavior - limit both models from achieving fully reliable counting and identity maintenance in real time. The report concludes that existing multi-object tracking methodologies are insufficient for tracking the swimming pattern of zebrafish. Enhanced detector training, more sophisticated motion models, and multi-camera or optimized lighting setups are promising directions for developing a deployable zebrafish monitoring solution.