Group Activity Recognition Based On Graph Attention Networks

Group activity recognition is a significant task in the domain of computer vision.The research has a lot of employment in real social life.The interaction of individuals and the spatial information of the group are two important cues for solving the problems of group activity recognition.The booming of deep learning methods drive the progress of the technology of group activity recognition.Graph convolution networks are drawing more and more attentions and playing more and more important roles in deep learning technology.In this thesis,Lastly,we mainly focus on following works:(1)We firstly introduce the background and practical meaning of group activity recognition.Then,we introduce the developing progress and present situation of individual action recognition 、 group activity recognition and graph convolution networks.Additionally,we present and analyze the basic methods and models separately based on hierarchical structure、interactive information、spatial information and multi-modal information.(2)Considering the importance of interaction and on the basis of graph attention network,We construct graph interactive attention network to fuse the interaction information of individuals in group and recognize the group activity.Then,experiments are made on Volleyball and Collective dataset to demonstrate the effectiveness of the graph interactive attention network.(3)To make full use of spatial information of group,We adopt action map to localize individuals and represent the action of individuals.Then,We construct graph spatial attention network to fuse the spatial information of group and recognize the group activity.Lastly,experiments are made on Volleyball dataset to demonstrate the effectiveness of the graph spatial attention network.The results on public Volleyball and Collective datasets demonstrate the effectiveness of our approaches.