학술논문
基于上下文感知空间坐标嵌入的时空图卷积网络 / Spatio-temporal Graph Convolutional Networks with Context-aware Spatial Coordinate Embedding
Document Type
Academic Journal
Author
Source
青岛大学学报（自然科学版） / Journal of Qingdao University(Natural Science Edition). 36(4):18-34
Subject
Language
Chinese
ISSN
1006-1037
Abstract
针对空间复杂的非欧几里得结构,图卷积网络不易通过欧氏距离构造输入图的问题,提出了上下文感知空间坐标嵌入的时空图卷积网络(STE-STA)模型,将空间背景和相关性明确地结合到模型中,并基于地理空间辅助任务学习、语义空间嵌入和动态图的时空注意力识别手势.首先从手骨架构造一个完全连接图,通过学习地理坐标的上下文感知向量编码,以及自我注意机制对节点特征和边缘进行自动学习;然后,与主任务并行预测数据中的空间自相关.实验结果表明,在DHG-14/28 数据集上,STE-STA模型识别率分别达到 92.40%与 87.85%,均高于目前最优模型;在 SHREC'17 数据集上,比时空图卷积网络(ST-GCN)分别高 0.60%和 0.10%.
For the complex non-Euclidean structure of space,graph convolutional network is not easy to construct the input graph through Euclidean distance,a context-aware spatial coordinate embeddingSpatio-Temporal Graph Convolutional Network(STE-STA)model was proposed,which explicitly combines spa-tial context and correlation into the model,and based on geospatial auxiliary task learning,semantic spa-tial embedding and dynamic graph spatio-temporal attention gesture recognition.Firstly,a fully connected graph was constructed from the hand skeleton,and the node features and edges were automatically learned by learning the context-aware vector encoding of geographic coordinates and the self-attention mechanism.Then,the spatial autocorrelation in the data was predicted in parallel with the main task.The experimen-tal results show that on the DHG-14/28 dataset,the recognition rate of the proposed algorithm reaches 92.40%and 87.85%,which are higher than the current optimal model.On the SHREC'17 dataset,it is 0.60%and 0.10%higher than Spatio-Temporal Graph Convolutional Network(ST-GCN).
For the complex non-Euclidean structure of space,graph convolutional network is not easy to construct the input graph through Euclidean distance,a context-aware spatial coordinate embeddingSpatio-Temporal Graph Convolutional Network(STE-STA)model was proposed,which explicitly combines spa-tial context and correlation into the model,and based on geospatial auxiliary task learning,semantic spa-tial embedding and dynamic graph spatio-temporal attention gesture recognition.Firstly,a fully connected graph was constructed from the hand skeleton,and the node features and edges were automatically learned by learning the context-aware vector encoding of geographic coordinates and the self-attention mechanism.Then,the spatial autocorrelation in the data was predicted in parallel with the main task.The experimen-tal results show that on the DHG-14/28 dataset,the recognition rate of the proposed algorithm reaches 92.40%and 87.85%,which are higher than the current optimal model.On the SHREC'17 dataset,it is 0.60%and 0.10%higher than Spatio-Temporal Graph Convolutional Network(ST-GCN).