【摘要】 在人体骨架行为识别中，图卷积网络可提取人体骨架拓扑结构来聚合特征信息。但现有方法既未有效关联骨架特征与拓扑关系，也忽略了不同视图下拓扑关系的变化性。为此，提出基于视图学习和通道特征拓扑融合的行为识别方法（VLCTF-GCN）。依据骨架的视图特征学习拓扑关系，为每个视图构建具有区分性的共享视图拓扑关系。在不同聚合程度上，结合视图与自适应拓扑关系，融合骨架通道特征与拓扑关系，使得拓扑结构能够自适应关联骨架特征，通过多尺度时间卷积提取不同时间长度的关节变化。在两个大型数据集的实验结果表明，所提方法性能优于现有方法。更多还原

【Abstract】 In skeleton-based human action recognition, graph convolutional networks can extract the topological structure of the human skeleton to aggregate feature information. However, existing methods not only fail to associate the skeletal features with the topological relationships effectively, but also overlook the variability of topological relationships under different views. To address this issue, an action recognition method based on view learning and channel feature topological fusion(VLCTF-GCN) was proposed. The topological relationships were learned according to the view features of the skeleton, and a discriminative shared view topological relationship was constructed for each view. At different aggregation levels, the view and the adaptive topological relationship were combined, and the skeletal channel features and the topological relationships were fused, enabling the topological structure to be adaptively associated with the skeletal features. Multi-scale temporal convolutions were used to extract the joint changes over different time lengths. Experimental results on two large-scale datasets demonstrate that the proposed method outperforms existing methods.更多还原