基于深度学习的人体动作识别方法  

Human motion recognition method based on deep learning

作　　者：沈西挺 于晟[1,2] 董瑶 董永峰 张泽伟[1,2] SHEN Xi-ting;YU Sheng;DONG Yao;DONG Yong-feng;ZHANG Ze-wei(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China;Hebei Provincial Key Laboratory of Big Data Computing,Hebei University of Technology,Tianjin 300401,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300401 [2]河北工业大学河北省大数据计算重点实验室,天津300401

出　　处：《计算机工程与设计》2020年第4期1153-1157,共5页Computer Engineering and Design

基　　金：天津市科技计划基金项目(14ZCDGSF00124);天津市自然科学基金项目(16JCYBJC15600)。

摘　　要：对基于机器视觉的人体动作识别的成果进行研究,为提高视频数据集中人体动作的识别率,提出一种改进的深度网络模型。采用稠密光流方法处理数据,结合二维卷积神经网络(2DCNN)、三维卷积神经网络(3DCNN)和长短期记忆神经网络(LSTM)对动作特征进行提取,利用Softmax分类器识别分类。通过KTH数据集进行实验对比验证,分析结果表明,改进模型相比其它已有模型具有更高的识别率,动作识别效果更优。The achievements of human motion recognition based on machine vision were deeply studied,to improve the recognition rate of human motion in video data sets,an improved depth network model was proposed.In this model,dense optical flow method was used to process data.Combining two-dimensional convolution neural network(2DCNN),three-dimensional convolution neural network(3DCNN)and long-term and short-term memory neural network(LSTM),motion features were extracted.Softmax classifier was used to recognize and classify.Experimental results on KTH data sets show that the improved model has higher recognition rate and better action recognition effects than other existing models.

关 键 词：人体动作识别 深度学习 KTH数据集 卷积神经网络 稠密光流 

分 类 号：TP24[自动化与计算机技术—检测技术与自动化装置]