Temporal Receptive Field Graph Convolutional Network for Skeleton-Based Action Recognition

Current methods for skeleton-based action recognition often prioritize spatial dimension dependencies. However, the temporal dimension dependency has not received sufficient attention. Moreover, while spatio-temporal graph convolution effectively captures the connections between adjacent frames, it fails to address the key issue of long-range spatio-temporal dependency modeling, which is crucial for providing more comprehensive action discrimination features. To tackle this problem, we propose a Temporal Receptive Field Graph Convolutional Network (TRFGCN). Within TRFGCN, we have designed a spatio-temporal receptive field module that aggregates features of adjacent frames and long-range spatio-temporal features in the skeleton sequence by generating receptive fields of varying sizes. Our evaluation on the large-scale NTU-RGB+D 60 dataset demonstrates that the proposed TRFGCN achieves superior performance in accurately classifying actions.