基于3D双流卷积神经网络和GRU网络的人体行为识别  

HUMAN ACTION RECOGNITION BASED ON 3D DUAL-FLOW CONVOLUTIONAL NEURAL NETWORK AND GRU NETWORK

作　　者：陈颖[1,2] 来兴雪[1] 周志全 秦晓宏[2] 池亚平[1,2] Chen Ying;Lai Xingxue;Zhou Zhiquan;Qin Xiaohong;Chi Yaping(College of Computer Science and Technology,Xidian University,Xi’an 710000,Shaanxi,China;Beijing Electronic Science and Technology Institute,Beijing 100070,China)

机构地区：[1]西安电子科技大学计算机科学与技术学院,陕西西安710000 [2]北京电子科技学院,北京100070

出　　处：《计算机应用与软件》2020年第5期164-168,218,共6页Computer Applications and Software

基　　金：国家重点研发计划项目(2018YF1004100)。

摘　　要：针对基于双流卷积神经网络的人体行为识别准确率不高,不能充分利用时间维度的信息问题,提出一种基于3D双流卷积和门控循环单元(GRU)网络的人体行为识别模型。将3D卷积神经网络引入到双流卷积神经网络中,在双流卷积神经网络的空间流和时间流中分别使用3D卷积神经网络提取视频的时空信息;融合3D双流卷积神经网络提取到的时空特征,形成有时间顺序的时空特征流;将时空特征流输入到具有记忆信息能力的GRU网络中递归学习时间维度的长时序列特征并利用线性SVM分类器进行人体行为识别。在行为识别数据集UCF101上的实验结果表明,该模型充分地利用了视频的时间维度信息,识别率为92.2%,优于其他人体行为识别算法。Aiming at the problem that the accuracy of human action recognition based on dual-flow convolutional neural network is not high and the information of time dimension cannot be fully utilized,we propose a human action recognition model based on 3D dual-flow convolutional and gated loop unit(GRU)network.It introduced the 3D convolutional neural network into the dual-flow convolutional neural network,and convolutional neural network was used to extract the spatial and temporal information of video in the spatial and temporal flow of the dual-flow convolutional neural network.Then,we fused the space-time features extracted by the 3D double-flow convolutional neural network to form a space-time feature flow with time sequence.The space-time feature flow was input into the GRU network with memory information capability to recursively learn the long-term sequence features of the time dimension,and the linear SVM classifier was used for human action recognition.The experimental results on action recognition data set UCF101 show that the model makes full use of the time dimension information of the video,and the recognition rate is 92.2%,which is superior to other human action recognition algorithms.

关 键 词：人体行为识别 3D卷积神经网络 双流卷积神经网络 门控循环单元 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]