Multi-Target Tracking Of Zebrafish Based On Improved HOG Features

The notable shoaling behaviour of zebrafish has garnered widespread interest among scientists.The accurate and rapid tracking of zebrafish shoals and obtaining behavioral data are the basis of the study of multi-zebrafish behavior.Additional,the 3D trajectory of zebrafish schools can enrich the shoaling information,it is crucial for relevant mechanistic exploration and predictions for ecosystems.Thus,we proposed a feature map suitable for identifying the back texture features of zebrafish,and we realized the accurate and stable tracking based on the identification of individuals.Finally,a stereo matching strategy of multi-view fusion is established to obtain the 3D trajectory.First,aiming at the problems of similar shape in zebrafish multi-target tracking,we proposed an improved HOG algorithm to calculate the stable back texture feature map of the zebrafish,combined with the SVM classification mechanism to achieve the identification of multi-zebrafish.Mainly,we classified the feature blocks based on the correlation of zebrafish back texture,also modified the way of feature output to improve the HOG feature suitable for describing the zebrafish back features.The identification rate of 30 zebrafish after 6 weeks was 60%,which shows the robustness of the algorithm.Additionally,we successfully identified the mice and fruit flies,which further verify the versatility of the algorithm.Then,as for the highly random trajectory of zebrafish,we proposed a series of stepwise enhanced data association strategies based on the identification.Thus,we tracked the multi-target of zebrafish in plane in a fully automated fashion,and obtained the final whole trajectories.The data association is divided into three levels: the initial tracklets acquisition based on the heuristic strategy,the tracklets lengthening based on timing correlation local classifier,the global classifier to connect the whole trajectory.Level enhancement,ensuring the tracking accuracy.The performance of the tracking algorithm was evaluated in 11 videos with different numbers and different sizes of zebrafish,and all of the tracking accuracy were over 95%,better than idTracker.Furthermore,it was well applied in zebrafish depression behavior experiment.Finally,we proposed the stereo matching strategy of view fusion to solve the occlusion problem in plane tracking.In particular,we simplified the surface area into a line area based on the skeleton extraction to deal the difficulty of target detection in the lateral view,and improved the efficiency of target detection.Further,based on the polar constraint,the vision fusion stereo matching method based on the tracklets was established.We used the uncrossed tracklets of top view and lateral view to connect the broken tracklets of one of the views,and achieved the three-dimensional stereo matching of the multi-target.The accuracy of the detection in the lateral view was improved by 6.84%,and the 3D trajectory was obtained at the same time.