基于图结构Transformer网络的骨架行为识别研究  

Research on Skeleton Action Recognition Based on Graph structured Transformer Networks

作　　者：刘子璇 俞建峰 钱陈豪 化春键 蒋毅 LIU Zixuan;YU Jianfeng;QIAN Chenhao;HUA Chunjian;JIANG Yi(School of Mechanical Engineering,Jiangnan University,Wuxi 214122,China;Jiangsu Key Laboratory of Advanced Food Manufacturing Equipmentand Technology,Wuxi 214122,China;School of Intelligent Manufacturing,Jiangnan University,Wuxi 214122,China)

机构地区：[1]江南大学机械工程学院,江苏无锡214122 [2]江苏省食品先进制造装备技术重点实验室,江苏无锡214122 [3]江南大学智能制造学院,江苏无锡214122

出　　处：《机械与电子》2025年第3期3-8,15,共7页Machinery & Electronics

基　　金：国家自然科学基金资助项目(51905215)。

摘　　要：针对目前人体骨架行为识别方法多用图卷积作为框架,缺少动作时序信息建模与时空特征融合能力的问题,提出一种基于Transformer算法的Actionformer模型。该模型采用分组注意力结构以加强局部特征的提取能力,并添加了空间信息嵌入与时序信息嵌入模块,以增强原始Transformer模型对空间和时间特征的提取。实验结果显示,Actionformer模型在NTU RGB+D数据集上的动作识别准确率较高,优于ST GCN和ST TR等基于图卷积和Transformer的传统模型。To address the limitations of current human skeleton based action recognition methods,which predominantly use graph convolutional frameworks and lack the ability to effectively model temporal information and fuse spatio temporal features,this paper proposes the Actionformer model based on the Transformer algorithm.This model employs a grouped attention structure to enhance the extraction of local features and incorporates spatial and temporal information embedding modules to improve the original Transformer model’s ability to capture spatial and temporal characteristics.Experimental results demonstrate that the Actionformer model achieves higher action recognition accuracy on the NTU RGB+D datasetthan traditional models based on graph convolution and Transformers,such as ST GCN and ST TR.

关 键 词：骨架动作识别 图结构Transformer 分层次注意力 骨架图结构嵌入 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]