Person-Frame Dynamic Feature Graph Network for Group Activity Recognition

Dynamic modeling of different dimensional features in video is the key element of group activity recognition. In the past years, a lot of work has been devoted to the modeling of character features, these methods have achieved good results, but most of them ignore that group activity is a continuous motion closely related to the scene, and underestimated the importance of the relationship between frames. This paper proposes a Person-Frame Dynamic Feature Graph Network to model group activity information from two levels: video frame level and individual level: Temporal Semantic sub-Graph (TSG) channel constructs temporal semantic relation subgraph for video frame features, and Person-level Dynamic Feature Map (PDFM) models personal dynamic characteristics. In addition, in order to alleviate the problem of slow training speed of group activity model, we use lightweight mobilenet-v2 as the backbone, and embed the Initial Feature Preprocessing Module (IFPM) in it to improve the training efficiency while maintaining the recognition accuracy. A lot of experiments have been done on this model with the most widely used dataset in the field of group activity recognition, and excellent results are obtained, which proves the effectiveness of the model.