FreTransLS: Frequency Transformer based large-scale group activity recognition model for sensor data

doi:10.1016/j.pmcj.2026.102179

2026-01-28

PRE

DOI：10.1016/j.pmcj.2026.102179

原文链接

原文求助

摘要

En 中文

In large-scale group activities, participants engage in a wider variety of actions, and the interactions among them become significantly more complex. This gives rise to challenges including synchronization and coordination analysis in group activity recognition. As a result, existing methods designed for recognizing small-scale group activities using sensor data often lead to inaccurate identification of dynamic patterns in large-scale settings. To address this issue, this paper proposes FreTransLS—a frequency Transformer-based model for large-scale group activity recognition using sensor data. FreTransLS introduces a novel approach for extracting time-frequency features in large-scale group activities. The approach integrates a spatio-temporal graph convolutional network (ST-GCN) module to capture spatio-temporal features within the group, along with a group location feature extraction (GLFE) module to acquire group location features. These two feature streams are fused to derive comprehensive time-domain representations of group activities. Furthermore, FreTransLS incorporates a frequency Transformer encoder built around a frequency attention mechanism. This encoder performs global analysis in the frequency domain to better model synchronization and coordination patterns in group activities. To enhance the generalization capability of the model, FreTransLS adopts a joint optimization strategy through complementary classification and reconstruction modules, which jointly refine the extracted time–frequency features. Experiments on two public datasets demonstrate that the proposed method effectively captures discriminative features from sensor data in large-scale group scenarios, leading to improved accuracy and robustness in group activity recognition.