Research And Application Of Zebrafish Trajectory Tracking Based On DeepSort

The animal multi-object tracking task currently is a popular research topic in the research of computer vision,and it has rich applications.Behavioral experiments of zebrafish have been widely used in developmental biology,gene function research,disease pathogenesis research,drug development,and other research fields.Accurate and rapid extracting of the three-dimensional trajectory of zebrafish is crucial for zebrafish behavior analysis,and can be used in related fields.The difficulty of 3D trajectory tracking of zebrafish lies in its similar appearance,occlusion,and rapid movement.In this paper,we propose a fish re-identification model suitable for zebrafish.On this basic,the zebrafish multi-target tracking in a single perspective is realized by combining DeepSort algorithm,and the three-dimensional trajectory tracking is realized by fusing multiple perspectives of cameras.Finally,a zebrafish behavior analysis platform is constructed.The main work of this thesis as follows:1.Aiming at the problem that the appearance and texture of zebrafish are extremely similar,and the target is easily lost during the tracking process,a zebrafish reidentification model based on metric learning is proposed.Inspired by person reidentification method,the matching of zebrafish individuals in different frames in the tracking process is regarded as an image retrieval subproblem.The similarity between different targets is calculated to achieve target matching.The results of zebrafish multitarget tracking experiments show that the model has excellent effect.2.Aiming at the problems of occlusion and rapid turning among individuals in the two-dimensional trajectory of zebrafish school in a single view,which leads to the tracking target to be easily lost,a multi-target tracking of zebrafish based on DeepSort algorithm is proposed.In this thesis,the two-dimensional trajectory tracking of the zebrafish is realized by data cascading,combining the individual appearance features obtained from the fish re-identification and the next frame motion information predicted by the Kalman filter algorithm.3.Aiming at the problem of data fusion of two perspectives in 3D trajectory,this thesis adopts the strategy of multi-view information fusion,combining the movement and appearance information between the top view and the side view,to achieve the trajectory matching of the two views,thus alleviate the problem of target loss in fish tracking,finally realizes the three-dimensional motion trajectory of zebrafish swimming.The results show that it can effectively obtain the three-dimensional trajectory of zebrafish.4.This thesis designs and implements a zebrafish behavior analysis system platform based on the needs of animal behavior analysis.The platform is based on the browser/server model and is developed using open source lightweight framework Springboot.This platform includes user management,trajectory visualization,behavior analysis,and data management functions.Users do not need to deploy and are easy to learn.They can directly visualize experimental results,which can greatly help the analysis of zebrafish motion behavior.