Scholarlink

<p>Spatiotemporal graph convolutional networks effectively capture spatial and temporal dependencies but suffer from high computational complexity and limited generalization, particularly in long-term and dynamic scenarios. To address these challenges, we propose temporal–spatial dynamic graph convolutional network (TSDGCN), which decouples temporal and spatial modeling. A SmallFeatureExtractor is introduced to compress feature dimensions and enhance local feature aggregation, improving early-stage representation. TSDGCN captures spatial dependencies through graph convolutions at each time step, while temporal dynamics are modeled using an independent temporal convolution module. Extensive experiments show that TSDGCN significantly improves long-term dependency modeling, adapts to dynamic graph structures, and reduces computational overhead. Ablation studies further validate the effectiveness of the decoupling strategy and feature extractor.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Temporal–spatial skeleton sequence recognition with self-supervised representation learning

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Hongwei Chen,
  • Shikai Kou,
  • Wei Wang,
  • Yuan Zhu,
  • Liya Xia

摘要

Spatiotemporal graph convolutional networks effectively capture spatial and temporal dependencies but suffer from high computational complexity and limited generalization, particularly in long-term and dynamic scenarios. To address these challenges, we propose temporal–spatial dynamic graph convolutional network (TSDGCN), which decouples temporal and spatial modeling. A SmallFeatureExtractor is introduced to compress feature dimensions and enhance local feature aggregation, improving early-stage representation. TSDGCN captures spatial dependencies through graph convolutions at each time step, while temporal dynamics are modeled using an independent temporal convolution module. Extensive experiments show that TSDGCN significantly improves long-term dependency modeling, adapts to dynamic graph structures, and reduces computational overhead. Ablation studies further validate the effectiveness of the decoupling strategy and feature extractor.