基于肢体子图的图卷积人体动作识别

Limb subgraph graph convolutional networks for skeleton-based action recognition

针对目前动作识别方法不能捕获肢体之间关联性关系特征问题,设计了基于肢体子图的图卷积动作识别方法。首先,为了捕获不同肢体的时空特征,将人体骨架图划分为相邻肢体有公共节点的4个肢体子图,分别表示相应的身体部位;其次,为了捕获不同肢体之间内在的关联性特征,设计了组合肢体子图,分别对肢体子图和组合肢体子图进行空间卷积,得到肢体的空间动作特征和空间关联性特征,通过公共节点聚合的空间特征,再对聚合特征进行时间卷积;最后,在训练阶段,为了优化模型,在每个时空单元中加入了残差结构,有效避免了因梯度消失导致子图特征值丢失的问题。在NTU-RGB+D和HDM05人体骨架数据集上验证了本模型的有效性。

To solve the problem that current motion recognition methods fail to capture the correlation features between limbs,this paper proposed a graph convolution action recognition method based on limb subgraph.First in order to capture the temporal and spatial characteristics of different limbs,this paper divided the human skeleton diagram into four limb subgraphs with common nodes of adjacent limbs,which represented the corresponding body parts,respectively.Second,for the sake of capturing the intrinsic relevance features between different limbs,acombined limb subgraph was designed,where the spatial motion features and spatial relevance features of limbs were obtained by spatial convolution of limb subgraph and combined limb subgraph respectively.The spatial features aggregated by common nodes are then temporal convolved.In the last training stage,to optimize the model,the residual structure was added to each spatial-temporal unit,which effectively avoided the loss of subgraph eigenvalues caused by gradient disappearance.The validity of the model was verified on NTU-RCB+D and HDM05 human skeleton data sets.