| The identification and trajectory tracking of biological individuals based on machine vision has always be a hotspot in the fields of biobehavioral research,agriculture and animal husbandry management,and wildlife protection.With the rise of deep learning technology,many novel technologies and methods have emerged in this field.The research on the behavioral characteristics of marine animals is an important basis for the management and protection of marine biological resources,and it is also a hot topic in the research on the behavioral response and adaptation of marine organisms under the background of global environmental change.At present,the use of machine vision methods to track the motion trajectories of marine animals mainly focuses on fish,and there are few studies on the behavioral tracking methods of large marine benthic animals.Because the behavior patterns of large Marine benthic animals are different from those of fish,many current behavior analysis methods are often not completely applicable.Charybdis japonica is an important economic crab species in the northern coast of my country,and its group behavior characteristics have important reference value for its resource multiplication and management.In this paper,Charybdis japonica is selected as the observation species,and research is carried out from different technical aspects such as individual identification,multi-target detection and tracking,and trajectory segment association,so as to provide reference for the establishment of behavioral research methods for large marine benthic animals.The main work of this paper is as follows:（1）Using the Resnet50 residual neural network as the feature extractor,the metric learning method was used to establish a Charybdis japonica individual re-identification model,meanwhile,the effects of loss function,feature extraction area,data enhancement strategy and feature vector dimension on model recognition accuracy are compared.The results show that the identification accuracy of the models established by the two different feature recognition regions of the whole body and the carapace has little difference,the feature vector dimension has a weak influence on the recognition accuracy of the model,and data enhancement can significantly improve the identification accuracy of the model.Combining with different application scenarios,the individual identification models established by Circle Loss and cross-entropy loss function are compared.It is found that in the same-ID application scenario,the Rank＿1 of the two model are similar,the former has a higher mean of average accuracy,however,the latter is more advantageous in cross-ID application scenarios.（2）The YOLOv5 model is used for Charybdis japonica target detection,the Charybdis japonica individual re-identification model is used for appearance feature extraction,and the Deep Sort algorithm is used to obtain the Charybdis japonica multi-target trajectory segment;and the network flow model and the minimum cost maximum flow algorithm are used to achieve trajectory segment association.to obtain accurate tracking of the trajectories of the Charybdis japonica population.The model evaluation results show that the target detection accuracy m AP of Charybdis japonica can reach 0.995,the multi-object tracking accuracy rate MOTA can reach 91.72%,and the maximum multi-object tracking precision rate MOTP is 20.79%.The trajectory segment association based on the network flow model effectively solves the ID switching problem caused by trajectory fracture in group tracking.The correction rate of track ID switching error is more than 90%.In this paper,the appearance feature extraction algorithm and network flow model of Charybdis japonica were established to effectively solve the trajectory breaking and ID switching problems of Charybdis japonica group tracking.It provides support for the study of Charybdis japonica colony behavior and the improvement of behavior analysis technology of large marine benthic animals. |
PDF Full Text Request