Group activity recognition requires a holistic understanding of individual actions, their spatial relationships, and the surrounding environment. Traditional methods that focus solely on isolated movements often fail to capture the complex inter-player and scene-level dependencies inherent in sports and crowd scenarios. In this research work, a model for group activity recognition is developed. The proposed model combines various contextual features through the integration of poses of individual actors in the scene with the pose-aligned spatial scene context for relational reasoning. Pose features of individual actors are extracted using mmPose, while the scene-level context is encoded through pose-conditioned spatial feature aggregation rather than explicit semantic segmentation. These pose and scene context features extracted are combined and used to construct Actor Relation Graphs (ARGs) using Zero Normalized Cross Correlation (ZNCC) which improves robustness to appearance and variations in illumination. Further, Graph Convolutional Networks (GCNs) are modelled using relationships between individual actors in a scene and their group activities. The proposed framework explicitly combines pose-level and scene-level contextual features into a single relational graph, in contrast to previous ARG-GCN approaches that mainly rely on appearance features. The model is evaluated on two benchmark datasets: the Collective Activity dataset (CAD) and the Volleyball dataset (VD). The model exhibits classification accuracies of 95.02% and 94.81% on CAD and VD, respectively. On a TITAN-XP GPU, the average time per video clip with 41 frames is approximately 0.2 s. The results show that the combination of pose and scene contexts features enhances graph-based relational learning and improves recognition accuracy.
Tejonidhi et al. (Fri,) studied this question.