Research On Group Behavior Recognition Based On Graph Convolutional Networks

Group behavior recognition is one of the research fields of computer vision behavior recognition,which has important academic and commercial value in intelligent monitoring,collective behavior analysis and sports video analysis.In recent years,with the rise of deep learning,Graph Convolutional Network(GCN)has gradually become a popular modeling method to describe the interaction relationship of group behavior recognition.Group activity recognition is an important task in computer vision,and the modeling of interaction between group members is its core technology.High complexity and information redundancy in relational reasoning are tough problems in complex scenarios when modeling its group interactions.In order to solve the problems,we propose a model of multigroup interactive relation.First of all,the scene information and individual information in the video frame as the initial features are extracted with CNN and Ro IAlign,and the coordinates of group members are used to divide the groups into two subgroups(for example,in the Volleyball dataset,create a serial ID for each member and then sort them with X coordinates of their bounding boxes,dividing them into two groups from left to right,that is,each contains six members).Secondly,the interaction models in two subgroups and the global scene group are inferred by Graph Convolutional Network,and further the key characters in each group are determined.Thirdly,we regard global relationship features as the real value,and cascade the characteristics of relation of the two subgroups as predicted value.The cross entropy loss function between the two groups is constructed to optimize the upper-level GCN network of group interaction,in order to ensure that the key players of the two subgroups and the global group key players can match successfully.Fourthly,after successful matching,the matched key figures within the two subgroups are regarded as the target nodes to establish inter-group relationship graph between these two subgroups,and it is then deduced by another GCN.Finally,the initial features are respectively fused with the intergroup and global interaction characteristics to obtain two group behavior recognition branches,and the final recognition result is obtained through their decision fusion.The experiment shows that the average accuracy rate is 93.1% in Volleyball dataset and 48.1% in NBA dataset.