Group activity recognition aims to identify behaviors involving multiple individuals.In real-world applications,group behavior is often treated as a hierarchical structure,which consists group,subgroups and individuals.Previous researches have been focused on modeling relationships between individuals,without in-depth relationship analysis between subgroups.Therefore,a novel hierarchical group activity recognition framework based on Multi-scale Sub-group Interaction Relationships(MSIR)is proposed,and an innovative multi-scale interaction features extraction method between subgroups is presented as specified below.A sub-group division module is implemented.It aggregates individuals with potential correlations based on their appearance features and spatial positions,then dynamically generates subgroups of different scales using semantic information.A sub-group interactive feature extraction module is developed to extract more discriminative subgroup features.It constructs interaction matrices between different subgroups and leverages the relational reasoning capabilities of graph neural networks.Compared with existing twelve methods on benchmark datasets for group behavior recognition,including volleyball and collective activity datasets,the methodology of this paper demonstrates superior performance.This research presents an easily extendable and adaptable group activity recognition framework,exhibiting strong generalization capabilities across different datasets.
维普
万方数据