GCN embedded with Polynomial Aggregation Function for Group Activity Recognition

Group activity detection is a crucial task for an automated analysis of human actions in many environments, such as surveillance or sport footage. Group activity recognition focuses on identifying the activities of one or more individuals through a sequence of observations on the conduct of the individuals. Numerous approaches have been proposed for the same using deep learning neural networks. In this paper, we focus on graphical convolutional neural network (GCN) which uses graphs as inputs, aggregate the vectors using aggregation functions and attempt to create models which can provide ample knowledge innate in human mobility. The paper aims to build a GCN method by adding the application of Polynomial Aggregation Function (PAF) to ensure the precision in investigating the activities is improved and further increases the accuracy of identification of human actions. The experiment is carried out using VGG16, GCN and GCN embedded with PAF on the Collective dataset. The experimental results obtained show that the whole group's actions as an entity is recognized and the proposed method provide better accuracy.