Group Activity Recognition Research Based On Graph Model And Deep Learning

Group activity recognition is a hot topic in computer vision.It has important application value in security monitoring,abnormal behavior detection,sports video analysis,etc.Regarding the interaction between group members,the previous modeling is relatively general and lacks the semantic description of the interaction between group members.However,in the interaction process of group members,there must be a relationship of cooperation or competition.Therefore,in order to solve the above Question,this article proposes a group behavior recognition model based on the interaction between cooperation and competition,and the model mainly contains two modules: weak supervision module and semi-supervised model.The weak supervision module is the rough judgment and reasoning of the automatic grouping and cooperation or competition of group members.First,the Faster RCNN is used to detect the members in the video sequence,and the IAP algorithm(Improved Affinity Propagation)is used to divide the group members into different clusters,and the Mobile Net network is used to extract the initial appearance features of each member,and then the full connection method is adopted.Construct an initial interaction diagram for each member of the cluster.In each cluster,there is not only a cooperative relationship between members,but also a competitive relationship.Therefore,in order to distinguish the two relationships between members and further complete the subdivision of the group,this article uses the sentiment analysis Bert network model and the two-category classification The processor analyzes the relationship between members in the same cluster to extract and identify whether different members are cooperative.Third,the graph convolutional network GCN is used to infer that the members of the cluster are cooperative or competitive,and the range of the cluster is continuously expanded and updated.The semi-supervised model is mainly to further supplement the member information.It uses GCN to realize the extraction of posture features in the video and realize the recognition of personal actions,and at the same time,the personal actions are labeled as supplementary features of the weakly supervised model.The features extracted by the semi-supervised model and the weakly-supervised module are merged through the Deep Layer Aggreation(DLA)fusion model,and the softmax classifier is used to recognize group behavior.In order to verify the effectiveness of the model in this paper,experiments were carried out on the CAD dataset and the NBA dataset,and the accuracy of 94.2% and 52.5% were achieved respectively..