基于行动片段补充生成器的异常行为检测方法  

Temporal Abnormal Behavior Localization Based on Action Fragment Supplement Generator

作　　者：赵春晖[1] 杨莹 宿南 ZHAO Chunhui;YANG Ying;SU Nan(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院

出　　处：《沈阳大学学报（自然科学版）》2019年第6期483-488,共6页Journal of Shenyang University：Natural Science

基　　金：国家自然科学基金资助项目（61405041,61571145）;中央高校基本科研业务费资助项目（3072019CFM0801,3072019CFQ0801）;哈尔滨市优秀学科带头人基金资助项目（RC2013XK009003)

摘　　要：为了解决监控视频数量的迅速增长给视频存储及分析带来的问题,提出了一种结合3D卷积网络与MIL(multiple instance learning)异常检测的方法,构造了一个异常行为片段补充生成器对提案网络的动作片段进行补充,并修改了分类网络的3D卷积网络结构,提升了分类网络的性能.根据MIL异常检测结果得分情况实现对边界检测结果的调整,自适应地控制输出结果的数量,在保证选择高分的异常行为检测结果的同时对多余部分进行筛选过滤,实现对监控视频的精细边界检测的目标.在UCF_crimes数据集上进行的实验表明,提出的异常行为边界检测方法与传统方法相比具有更好的检测效果.In order to solve the problems of video storage and analysis caused by the rapid growth of the number of surveillance video,a method combining 3D convolution network with MIL(multiple instance learning)anomaly detection is proposed.An abnormal behavior fragment supplement generator is constructed to supplement the action fragments of the proposed network,and the 3D convolution network structure of the classification network is modified to improve the performance of the classification network.According to the MIL abnormal detection results score,the adjustment of boundary test results is realized,the quantity of output results is controlled adaptively,and the excess part is filtered and filtered while the abnormal behavior test results of high scores are guaranteed to be selected,so as to achieve the goal of fine boundary detection of surveillance video.Experiments on the UCF-crimes data set show that the proposed abnormal behavior boundary detection method has a better detection effect than the traditional method.

关 键 词：监控视频 多实例学习 3D卷积 异常行为 边界检测 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]