文献类型：期刊文献

中文题名：基于深度神经网络的多视角人体动作识别

英文题名：Multi-view Human Action Recognition Based on Deep Neural Network

作者：赵瑛[1,2];陆耀[1];张健[3];梁启弟[3];龙炜[1]

第一作者：赵瑛

机构：[1]北京理工大学智能信息技术北京市重点实验室,北京100081;[2]北京联合大学师范学院,北京100011;[3]中南大学计算机学院,长沙410083

第一机构：北京理工大学智能信息技术北京市重点实验室,北京100081

年份：2021

卷号：33

期号：5

起止页码：1019-1030

中文期刊名：系统仿真学报

外文期刊名：Journal of System Simulation

收录：CSTPCD;;Scopus;北大核心:【北大核心2020】；CSCD:【CSCD2021_2022】；

基金：国家自然科学基金(61273273);国家重点研发计划(2017YFC0112001)。

语种：中文

中文关键词：多视角;人体动作识别;卷积神经网络;上下文注意力;序列匹配

外文关键词：multi-view;human action recognition;convolutional neural network;context attention;sequence matching

摘要：为提高多视角人体动作识别的精度,提出了一种新的深度神经网络模型——CNN+CA(Convolutional Neural Networkplus Context Attention)模型和一种基于序列匹配的识别方法。利用卷积神经网络自动学习出多视角融合特征;引入上下文注意力模块自动突出特征中有利于识别的区域,进一步提高特征的判别力;通过基于序列匹配的方法实现人体动作识别。在IXMAS数据集和i3DPost数据集上的实验结果表明,所提方法在识别精度上超过了其他同类方法。
A novel deep neural network named CNN+CA(Convolutional Neural Network plus Context Attention)model is constructed and a new recognition algorithm based on sequence matching is presented to improve the recognition accuracy of MVHAR(Multi-view Human Action Recognition).A CNN(Convolutional Neural Network)is designed to automatically learn multi-view fusion features;the CA(Context Attention)module is introduced to selectively focus on the parts of the features that are relevant for the recognition task;the proposed recognition algorithm based on sequence matching is used to realize MVHAR.The experimental results on the IXMAS dataset and the i3 DPost dataset demonstrate that the recognition accuracy of the proposed method is higher than those of the state-of-the-art MVHAR methods.