Design And Implementation Of A Group Behavior Recognition System Based On Dual Graph Convolutional Network

Group activity recognition is one of the popular research fields in multimedia and computer vision,it is widely utilized in many scenes such as video surveillance analysis and sports video analysis.Compared with recognizing the actions of individuals in images,group activity recognition in videos not only needs to consider about temporal changes,but also integrate the feature of multiple individuals in the scene.In recent years,group activity recognition has undergone a shift from manual feature extraction to deep learning methods,and has made leapfrog progress with the development of deep learning.A number of works have shown that learning the relationship among actors in the scene can effectively improve the accuracy of recognition.However,the existing methods have too much coverage for the learning of relationships among actors and the learned relationship is relatively single.For existing problems in group activity recognition,this thesis designs a group activity recognition algorithm based on Dual Graph Convolutional Network and a group activity recognition algorithm based on Dual Graph Convolutional Network with Motion Weights,then implements a group activity recognition system.The specific works of this thesis are as follows:(1)This thesis proposes a group activity recognition method based on a Dual Graph Convolutional Network.This model is mainly composed of two modules,which are the Pose Guided Pruning Module(PGP)and the Dual Graph Convolutional Network(Dual-GCN)module.The PGP module uses the actor’s skeletal point features to describe the posture of the human body and obtain the actor’s attention range.By analyzing the range of attention,the key actors are obtained,and the key regions are drawn with them as the center of the circles.The Dual-GCN module uses a two-layer Graph Convolutional Network to learn the distance among actors in the key region and the attention coverage relationship.The experimental results show that the group activity recognition method based on Dual Graph Convolutional Network proposed in this thesis can effectively improve the accuracy of group activity recognition.(2)This thesis proposes a Dual Graph Convolutional Network based on Motion Weights,which further improves the accuracy of group activity recognition.We refer to the method of quantifying the activity of the actor in the previous work,propose a method to score the motion of the actors in the scene and generate the motion weights.On the basis of the Dual Graph Convolutional Network model,the weight of the movement is added to describe the relationship among actors more accurately.The experimental results show that the Dual Graph Convolutional network for group activity recognition algorithm with Motion Weights can further improve the accuracy of recognition.(3)This thesis designs a group activity recognition system based on Dual Graph Convolutional Network for group activity recognition.Specifically,it is divided into two parts: the foreground system and the background system.We implemented the functions of the system according to the requirements,and applied the group activity recognition algorithm based on the Dual Graph Convolutional Network proposed in this thesis to the system,which can realize the recognition of video activity at a fast rate.The system can also save and query the previous recognition results.