基于时序拓扑非共享图卷积和多尺度时间卷积的骨架行为识别

Temporal Topology Unshared Graph Convolution and Multiscale Temporal Convolution for Skeleton-based Action Recognition

  • 摘要: 针对基于图卷积的骨架行为识别中不同骨架帧共享同一空间拓扑及采用单一尺度时间卷积建模时域特征的缺点,提出一种基于时序拓扑非共享图卷积和多尺度时间卷积的行为识别方法。首先,在空间建模中根据输入样本计算每帧骨架上关节之间的关系,为每帧骨架构建独立的空间拓扑;其次,在时间建模中采用具有5条分支的多尺度时间卷积模块提取不同时间尺度上的行为特征;最后,结合时序拓扑非共享图卷积和多尺度时间卷积模块构建了一个用于骨架行为识别的时空图卷积网络。在NTU RGB+D、NTU RGB+D 120和Northwestern-UCLA数据集上的实验结果表明,所提方法在识别准确率和模型复杂度方面优于主流的行为识别方法。

     

    Abstract: In skeleton action recognition based on graph convolution, the different skeleton frames share the same spatial topology, and the temporal feature model employs single-scale temporal convolution. We address these issues and propose an action recognition methodology based on temporal topology unshared graph convolution and multiscale temporal convolution. First, in spatial modeling, we calculate the joint relationship of each frame according to the input samples to establish the independent spatial topology for each skeleton frame. Second, we use the multiscale temporal convolution module with five branches in the temporal modeling to extract action features on different time scales. Finally, we propose a spatiotemporal graph convolutional network for the skeleton action recognition by combining the temporal topology unshared graph convolution and multiscale temporal convolution modules. We carry out comparative experiments on NTU RGB+D, NTU RGB+D 120, and Northwestern-UCLA datasets. The results show that the proposed method has better recognition accuracy with lower model complexity than the current main action recognition methods.

     

/

返回文章
返回