DGCMam: Fusing Distance Graph Convolution and Mamba for Skeleton-based Action Recognition

Most skeleton-based action recognition algorithms extract spatio-temporal features from skeleton data using Graph Convolutional Networks (GCNs) and Transformers. However, typical GCNs use static learnable parameters to aggregate joint features, which prevents the model from dynamically adjusting aggregation weights based on different motion patterns. Although the Transformer can dynamically aggregate features by the self-attention mechanism, it has high computational complexity when processing long sequences. To address these limitations, we propose DGCMam, a model fusing distance graph convolution and Mamba for skeletonbased action recognition. DGCMam uses the Distance Graph Convolution (DGC) to extract features, where learnable distance parameters are applied to adjust the aggregation weights of joints based on spatial distances. Moreover, DGCMam effectively extracts global spatio-temporal features through the alternating-direction Mamba (AD-Mamba) structure, achieving lower computational complexity than bidirectional Mamba and Transformers for long sequence data. A gated fusion mechanism is also employed to enhance high-level features by selectively adding initial features. Experiments on the NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets show that DGCMam achieves the state-of-the-art performance with lower inference time.