基于自动编码器和神经网络的人体运动识别  

Human action recognition using auto-encoder and neural network

作　　者：司阳 肖秦琨[1] 李兴 Si Yang;Xiao Qinkun;Li Xing(Department of Electronic Information Engineering, Xi＇an Technological University, Xi＇an 710021,China)

机构地区：[1]西安工业大学电子信息工程学院

出　　处：《国外电子测量技术》2018年第1期78-84,共7页Foreign Electronic Measurement Technology

基　　金：国家自然科学基金项目(60972095,61271362,61671362);陕西省自然科学基金项目(2017JM6041)资助

摘　　要：当前传统的人体运动识别是基于卷积神经网络（CNN）的方法,但存在原始数据维数高,利用CPU训练时间长及硬件要求高的缺点。针对以上问题,提出一种由自动编码器与模式识别神经网络（PRNN）组成的识别人体运动的深度神经网络模型。算法分为系统学习阶段和动作识别阶段。在系统学习阶段,首先得到每帧的人体轮廓,构建二进制重叠图像作为训练数据,并训练一个自动编码器来提取动作特征;其次,利用所得到的特征通过监督学习训练PRNN;最后建立新的深度神经网络,通过微调获得最佳性能。在动作识别阶段,人体的运动行为序列首先被翻译成二进制重叠图像,然后使用APRNN进行识别。测试结果表明,这种方法具有很好的性能。Currently,traditional body movement identification is based on the convolution neural network（CNN）method,but the original data dimension is high,and the CPU training time and hardware requirements are high.so developed a new deep neural network model to identify human action that was composed of an auto-encoder and a pattern recognition neural network（PRNN）.Our approach was divided into two parts：a system learning stage and an action recognition stage.In the system learning stage,first we secured human body outlines for each image frame,and combined the outlines to build an overlay of binary images to use as training data.Based on deep neural network learning,an auto-encoder was trained to extract action features.Next,we used supervised learning to train a PRNN on the obtained features.Last,we combined the auto-encoder with the PRNN to build a new deep neural network called the APRNN.Using fine tuning,the APRNN achieved optimal performance.In the action recognition stage of our approach,human action sequences were translated into binary overlay images,and the ARPNN was used to identify the actions.Test results showed our method had better performance than existing approaches.

关 键 词：动作识别 自动编码器 模式识别神经网络 深度神经网络 二进制重叠图像 

分 类 号：TN911.73[电子电信—通信与信息系统]