Tracking Multiple Zebrafish Larvae Using YOLOv5 and DeepSORT

Accurate tracking of zebrafish larva movements is essential to examine their dynamic behaviors in biomedical and pharmaceutical applications. However, the characteristic “burst” and sleep-like stationary movements of zebrafish will cause inconsistency in the detection and tracking, which hinders the observation of their identities and trajectories. To address these problems, this paper develops an accurate and reliable tracking system for multiple zebrafish larvae based on the current state-of-the-art detection technology YOLOv5 and multi-target tracking technology DeepSORT. The detection-based tracking system divided into two parts: detection and tracking. In the detection stage, the zebrafish larvae’s head position detected by using the trained YOLOv5 model. In the DeepSORT tracking phase, a Cascade matching algorithm utilized to match the identification of zebrafish larvae, which utilizes information on movement and appearance of larvae. On the one hand, the Mahalanobis distance used to evaluate the predicted value by the Kalman filter and the detected value by the YOLOv5 model. On the other hand, the appearance features model of the fish head trained and utilizes Cosine distance to evaluate between the stored values of the appearance features and the current frame appearance information. The proposed tracking system has been evaluated by using CLEAR MOT Metrics on five groups of zebrafish larvae videos under various complex imaging conditions. The results showed that the system had good performance in reducing the ID switch problem in complex environments, and the tracking accuracy could be up to 88.8%.