结合语义概念和双流特征模型的复杂事件检测  

Complex event detection based on combination of semantic concept and two-stream feature model

作　　者：张建明[1] 黄伟康 詹永照[1] ZHANG Jianming;HUANG Weikang;ZHAN Yongzhao(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang,Jiangsu 212013,China)

机构地区：[1]江苏大学计算机科学与通信工程学院

出　　处：《江苏大学学报（自然科学版）》2020年第1期60-66,共7页Journal of Jiangsu University：Natural Science Edition

基　　金：国家自然科学基金资助项目(61672268);江苏省重点研发计划项目(BE2015137)

摘　　要：针对已有复杂视频事件检测方法未能有效利用语义概念信息的问题,提出了结合语义概念和双流特征模型的复杂事件检测方法.该方法采用动作检测器和对象概念检测器,得到动态概念和静态概念.提出针对任务的优选概念子集生成方法,并以此构建基于优选概念子集的视频事件检测器.同时构建光流图像和空间流序列的双流特征卷积神经网络模型加LSTM的事件分析表达模型,进而将两流事件分析结果进行融合分类检测.最后将基于语义概念的事件分类分析结果和基于双流模型的事件分类分析结果进行决策融合,最终检测出复杂事件.在典型的复杂事件数据集上将所提算法与相关算法进行了对比试验.结果表明,所提的方法有了实质性的改进,准确率达到了81.1%,相比于最优算法提高了5.7%.To solve the problem that the traditional complex video event detection methods could not effectively use the semantic concept information,the complex event detection method was proposed based on the combination of semantic concept and two-stream feature model.The motion detector and the object concept detector were employed to obtain dynamic concepts and static concepts.A method of generating optimal concept subsets for tasks was proposed to construct the video event detector according to the preferred concept subset.Combined with the two-stream feature and LSTM,the results of event analysis and expression model were fused and classified.The event classification analysis results respectively based on the semantic concept and the dual-stream model were merged with decision-making to detect complex events.The proposed algorithm was compared with the related algorithms on the typical complex event dataset.The results show that the proposed method is modified substantially with accuracy rate of 81.1%,which is 5.7% higher than that of the optimal algorithm.

关 键 词：视频内容理解 语义概念 深度学习 事件检测 视频分析 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]