Deep Learning-Based Group Activity Recognition: In-Depth Analysis Based on Videos

Group activity recognition (GAR) has several practical applications such as analysis of sports videos, surveillance and security, crowd management, public safety, etc. Researchers are focusing more on group activity recognition using videos. A state-of-the-art review is given in this paper for Group activity recognition in videos. First the importance of Group activity recognition is explored, then a survey on group activity recognition methods is given which includes several categories of group activity recognition like actor relation based, attention based, Recurrent Neural Network (RNN) based, skeleton based and graph based. Several datasets which are commonly used in group activity recognition are discussed in detail. As group activity recognition performs classification of different activities, performance metrics of classification are elaborated. Finally, through several challenges conceivable directions for upcoming researchers are summarized.