基于多级图像序列和卷积神经网络的人体行为识别  

Multi-level image sequences and convolutional neural networks based human action recognition method

作　　者：马淼[1] 李贻斌[1] 

机构地区：[1]山东大学控制科学与工程学院,济南250061

出　　处：《吉林大学学报（工学版）》2017年第4期1244-1252,共9页Journal of Jilin University:Engineering and Technology Edition

基　　金："863"国家高技术研究发展计划项目(2015AA042201);国家自然科学基金项目(61233014)

摘　　要：首先,构造出能获得更丰富人体行为信息的四级图像序列结构,并分别用卷积神经网络进行处理,从而得到包含表观、运动、前景和背景信息的特征。然后,提出了一种对视频中行为进行分解的方法,将完整行为分解为由粗略到细致的子行为,从而得到更细致的人体行为描述,获取到更具代表性的行为特征。最后,通过两个行为数据集上的验证及对比实验证明了该方法可有效提高行为识别的准确度。A multi-level image sequences and convolutional neural networks human action recognition method is proposed.First,a four-level image sequence structure is constructed,which is able to obtain richer information of human actions.Then the four-level image sequences are processed by convolutional neural networks.This structure is able to use appearance,motion,foreground and background information more sufficiently.Besides,a decomposition method of video sequence is proposed,which is able to acquire more detailed human activity information.This method decomposes each level sequence into sub-sequences,and represents actions from coarse to fine,thus,achieving more representative human activity features.The efficiency of the proposed method is verified by two challenging human action databases.The experiment results show that the proposed method improves the action recognition accuracy efficiently.

关 键 词：人工智能 行为识别 视频理解 卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]