面向人机交互的通道注意力位移图神经网络  

Channel Attention Shift Graph Neural Network for Human-computer Interaction

作　　者：易思恒 陈永辉[1] 王赋攀[1] 蔡婷 YI Si-heng;CHEN Yong-hui;WANG Fu-pan;CAI Ting(College of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621000,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621000

出　　处：《小型微型计算机系统》2022年第3期604-610,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61802320,61872304)资助;西南科技大学博士基金项目(18zx7105,19zx7144)资助。

摘　　要：在人机交互动作识别领域中,基于深度学习的动作识别方法比传统的手工特征提取方法准确率更高.为了解决基于深度学习的动作识别方法在实时人机交互的实际应用问题,本文设计并创建了交互动作数据集(IA RGB-D),用于深度学习方法的人体动作识别研究.将IA RGB-D用于多种神经网络的训练和测试,测试结果准确率均在95%以上,验证了数据集的正确性和有效性.为保障对采集动作的实时识别正确率,本文提出了一种基于高效通道注意力的位移图神经网络(ASGCN),将高效通道注意力模块引入位移图卷积神经网络(Shift-GCN),增强其在通道特征上的提取能力.实验证明,ASGCN比Shift GCN准确率更高,提高了复杂动作的识别率,并且与传统的手工特征提取方法对比,识别效率接近但是准确率大幅提升.In the field of human-computer interaction action recognition,the method of action recognition based on deep-learning has higher accuracy than than the methods of traditional handcrafted features.In order to solve the practical application of action recognition methods based on deep learning in real-time human-computer interaction,an interactive action RGB-D data set(IA RGB-D)is proposed.This data set is used for deep learning methods of human action recognition research.IA RGB-D is used to train a variety of neural networks,and the accuracy of the test results was all above 95%,which verifies the correctness and effectiveness of the data set.In order to ensure the correct rate of real-time recognition of collected actions,an attention-based Shift graph neural network(ASGCN)is proposed,which integrates efficient channel attention module into Shift graph neural network(Shift-GCN)to enhance its ability to extract channel features.Experiments have proved that ASGCN has a higher accuracy rate than Shift GCN and improves the recognition rate of complex actions.Compared with the methods of traditional handcrafted features,the recognition efficiency is close but the accuracy rate is greatly improved.

关 键 词：人体动作识别 图卷积神经网络 人机交互动作数据集 人机交互动作识别 骨骼关节点数据 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]