基于骨架动作识别的分块广义图卷积网络  

Block Generalized Graph Convolutional Network Based on Skeleton Action Recognition

作　　者：杨超 丁文文[1] 邓淦森 YANG Chao;DING Wenwen;DENG Gansen(School of Mathematical Sciences,Huaibei Normal University,235000,Huaibei,AnHui,China)

机构地区：[1]淮北师范大学数学科学学院,安徽淮北235000

出　　处：《淮北师范大学学报（自然科学版）》2023年第3期66-70,共5页Journal of Huaibei Normal University：Natural Sciences

基　　金：国家自然科学基金项目(62171342);安徽省自然科学基金项目(1908085MF186)。

摘　　要：针对传统图卷积网络易忽略骨架图的结构性问题,设计一种基于骨架分块和构造广义图卷积网络模型。首先,通过谱图理论捕获时空变化。其次,提出一种用于骨架动作识别的分块广义图卷积网络,利用时空图来自然地表示人体动作序列。特别是对人体骨架进行空间划分,获取人体部分之间的关系。构造广义图,获取时间维度上的关系。实验结果表明,PG-GCN模型在NTU RGB+D 60数据集的CS和CV中的识别率分别为88.9%、95.2%。与较为先进的方法相比,在CS与CV的标准上分别提升4.1%、2.8%,证明该方法具有一定的先进性。Aiming at the problem that the traditional graph convolutional network is easy to ignore the structure of the skeleton graph,A generalized graph convolutional network model was designed based on skeleton block and construction.Firstly,the spatio-temporal variation is captured by spectral graph theory.Secondly,a block generalized graph convolutional network for skeletal action recognition is proposed,which uses spatio-temporal graphs to naturally represent human action sequences.In particular,the human skeleton is spatially partitioned to obtain the relationship between human parts.The generalized graph is constructed to capture the relationship in the time dimension.The experimental results show that the recognition rates of PG-GCN model in CS and CV of NTU RGB+D 60 dataset are 88.9%and 95.2%,respectively.Compared with the more advanced methods,the proposed method improves CS and CV standards by 4.1%and 2.8%,respectively.This method is proved to be advanced.

关 键 词：动作识别 分块广义图 图卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]