基于图注意与有向图神经网络的人体动作识别

Human Action Recognition Method Based on Graph Attention Network and Directed Graph Neural Network

基于人体骨骼数据的图卷积神经网络不易受背景环境噪声影响且鲁棒性较强,已成为现阶段人体动作识别领域的研究重点,但该网络对同阶邻域中不同邻域赋予相同权值,限制了其捕捉空间信息相关性的能力。为此,引入图注意网络加权和求和相邻节点的特征,允许每个节点根据其相邻特征分配不同权重,以增强特征提取和学习能力。同时,为解决将骨架表示为无向图时只能确定相邻节点或边之间的关系,从而限制了捕获节点或边之间依赖关系能力这一问题。引入有向图卷积,利用一阶和二阶相邻节点的特征信息进行图卷积,既保留了有向图的方向性特征,又扩展了图卷积的感知域,从而能够提取更多特征。实验表明,所提方法能有效提升动作识别的精度。

The graph convolutional neural network based on human skeleton data is not easily affected by background environmental noise and has strong robustness,which has become a research focus in the field of human action recognition at present.However,this network assigns the same weight to different neighborhoods in the same order,which limits its ability to capture spatial information correlations.To this end,a graph attention network weighted sum is introduced to sum the features of adjacent nodes,allowing each node to assign different weights based on its adjacent features to enhance feature extraction and learning effectiveness.At the same time,in order to solve the problem of representing the skeleton as an undirected graph,only the relationship between adjacent nodes or edges can be determined,which limits the ability to capture dependency relationships between nodes or edges.Introducing directed graph convolution,utilizing the feature information of first-order and second-order adjacent nodes for graph convolution,not only preserves the directional features of the directed graph,but also expands the perceptual domain of graph convolution,thereby extracting more features.The experiment shows that the proposed method can effectively improve the accuracy of action recognition.