基于多层卷积神经网络特征和双向长短时记忆单元的行为识别（英文）

发布时间：2018-01-16 04:20

本文关键词：基于多层卷积神经网络特征和双向长短时记忆单元的行为识别（英文）　出处：《控制理论与应用》2017年06期 　论文类型：期刊论文

【摘要】：鲁棒的视频行为识别由于其复杂性成为了一项极具挑战的任务.如何有效提取鲁棒的时空特征成为解决问题的关键.在本文中,提出使用双向长短时记忆单元(Bi-LSTM)作为主要框架去捕获视频序列的双向时空特征.首先,为了增强特征表达,使用多层的卷积神经网络特征代替传统的手工特征.多层卷积特征融合了低层形状信息和高层语义信息,能够捕获丰富的空间信息.然后,将提取到的卷积特征输入Bi-LSTM,Bi-LSTM包含两个不同方向的LSTM层.前向层从前向后捕获视频演变,后向层反方向建模视频演变.最后两个方向的演变表达融合到Softmax中,得到最后的分类结果.在UCF101和HMDB51数据集上的实验结果显示本文的方法在行为识别上可以取得较好的性能.
[Abstract]:In this paper , we propose the use of bi - directional long - term memory cell ( Bi - LSTM ) as the main frame to capture the two - way space - time characteristics of the video sequence . Firstly , in order to enhance the feature expression , the multi - layer convolution neural network feature is used instead of the traditional manual feature . The multi - layer convolution feature combines the low - level shape information and the high - level semantic information to capture the rich spatial information .

【作者单位】：苏州大学计算机科学与技术学院;吉林大学符号计算与知识工程教育部重点实验室;软件新技术与产业化协同创新中心;常熟理工学院计算机科学与工程学院;
【基金】：Supported by National Natural Science Foundation of China(61170124,61272258,61301299,61272005,61572085) Provincial Natural Science Foundation of Jiangsu(BK20151254,BK20151260) Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education,Jilin University(93K172016K08) a Prospective Joint Research Projects from Joint Innovation and Research Foundation of Jiangsu Province(BY2014-05914) Collaborative Innovation Center of Novel Software Technology and Industrialization
【分类号】：TP183;TP391.41
【正文快照】： 1 IntroductionThe first challenge is the selection of powerful fea-Action recognition[1]is one of the most active areasture representations.In the last several decades,manydue to its wide applications,such as video surveillance,hand-crafted features are

【相似文献】