Zebrafish is an excellent model organism, which has been widely used in the fields of biological experiments, drug screening, and swarm intelligence. In recent years, there are a large number of techniques for tracking of zebrafish involved in the study of behaviors, which makes it attack much attention of scientists from many fields. Multi-target tracking of zebrafish is still facing many challenges. The high mobility and uncertainty make it difficult to predict its motion; the similar appearances and texture features make it difficult to establish an appearance model; it is even hard to link the trajectories because of the frequent occlusion. In this paper, we use particle filter to approximate the uncertainty of the motion. Firstly, by analyzing the motion characteristics of zebrafish, we establish an efficient hybrid motion model to predict its positions; then we establish an appearance model based on the predicted positions to predict the postures of every targets, meanwhile weigh the particles by comparing the difference of predicted pose and observation pose ; finally, we get the optimal position of single zebrafish through the weighted position, and use the joint particle filter to process trajectory linking of multiple zebrafish.