Objectives To address the challenges posed by special operational scenarios and limited public data in carrier aircraft deck operations, this study proposes a recognition method based on multi-dimensional features.

Methods First, key points such as channel boundaries and static obstacles are accurately selected to represent the environmental information. Interactions between dynamic individuals and static environmental objects are modelled using graph convolutional networks to explore their underlying connections of operational object interactions. Then, a multi-scale spatio-temporal feature extraction module is designed, incorporating a dilated attention mechanism that captures key individual interactions at both global and local levels by applying different dilation rates. At the same time, temporal convolutional networks (TCN) combined with the attention mechanism are employed to extract temporal interaction feature, efficiently capturing dynamic relationships across both long and short sequences. Finally, the multi-scale spatio-temporal feature extraction module is stacked multiple times to adaptively extract multi-dimensional features, thereby improving the recognition accuracy of carrier aircraft deck operations.

Results Experimental conducted on a self-constructed dataset featuring multi-perspective carrier deck operation scenarios involving heterogeneous objects demonstrate that the proposed method significantly outperforms existing group activity recognition models such as ARG, DIN, AT, and GroupFormer, achieving an accuracy of 97.8%.

Conclusions This study provides a valuable reference for the high-accuracy recognition of carrier aircraft deck operations.