摘要
针对卷积操作受到遍历规则的限制,只能提取单个骨骼节点的特征信息,不能对相邻节点之间的有效特征信息进行融合,导致表达能力有限的问题,提出了一种基于特征位移模块的手势识别神经网络。该网络采用常规时空图卷积神经网络的架构,并将常规时空卷积模块替换为特征位移模块,实现相邻节点特征信息之间的融合。利用特征位移模块对位移信道进行重新排序,实现提取骨骼节点的全局化特征信息,进一步完成对手势信息的高效准确分类。并在公开数据集DHG-14/28和FPHA上验证该特征位移模块,在14类、28类和FPHA手势数据集的分类准确度分别达到了95.11%、93.01%和92.67%。实验结果表明,该网络模型能够更好更有效的挖掘全局特征信息,在常见的手势识别数据集上达到了优秀的性能。
The convolutional operation is constrained by traversal rules,limiting the extraction of feature information from individual skeletal nodes and preventing effective fusion of feature information between adjacent nodes,resulting in limited expressive power.In response to this issue,a gesture recognition neural network based on a Feature Displacement Module is proposed.This network adopts the architecture of conventional spatiotemporal graph convolutional neural networks and replaces the conventional spatiotemporal convolution module with the Feature Displacement Module to achieve fusion of feature information between adjacent nodes.By reordering the displacement channels through the Feature Displacement Module,global feature information of skeletal nodes is extracted,further enabling efficient and accurate classification of gesture information.The Feature Displacement Module is validated on the public dataset DHG-14/28 and FPHA,achieving classification accuracies of 95.11%,93.01% and 92.67% for 14-class,28-class and FPHA gesture datasets.The experimental results demonstrate that this network model can better and more effectively mine global feature information,achieving excellent performance on common gesture recognition datasets.
作者
刘翔
刘新妹
李传坤
张晋钊
Liu Xiang;Liu Xinmei;Li Chuankun;Zhang Jinzhao(School of Information and Communication Engineering,North University of China,Taiyuan 030051,China;State Key Laboratory of Dynamic Testing Technology,North University of China,Taiyuan 030051,China)
出处
《电子测量技术》
北大核心
2024年第8期141-147,共7页
Electronic Measurement Technology
基金
山西省回国留学人员科研项目(2017-090)
山西省重点研发项目(201903D121058)资助。
关键词
手势识别
卷积神经网络
特征位移
图卷积神经网络
深度学习
gesture recognition
convolutional neural network
feature shift
graph convolutional networks
deep learning
