端对端SSD实时视频监控异常目标检测与定位算法  

End-to-end SSD real-time video surveillance anomaly detection and location

作　　者：胡正平[1,2] 张乐 李淑芳[1,2] 赵梦瑶 HU Zhengping;ZHANG Le;LI Shufang;ZHAO Mengyao(School of Information and Engineering,Yanshan University,Qinhuangdao,Hebei 066004,China;Hebei Key Laboratory of Information Transmission and Signal Processing,Qinhuangdao,Hebei 066004,China)

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004 [2]燕山大学河北省信息传输与信号处理重点实验室,河北秦皇岛066004

出　　处：《燕山大学学报》2020年第5期493-501,共9页Journal of Yanshan University

基　　金：国家自然科学基金资助项目(61771420);河北省自然科学基金资助项目(F2016203422)。

摘　　要：为实现复杂场景下端对端实时视频监控异常目标检测与定位,借鉴目标检测思路,提出端对端SSD实时视频监控异常目标检测与定位算法。本算法在卷积神经网络6个不同尺度卷积特征图上采用2组3×3卷积核设置目标预选框得到异常分类及更加准确完整的异常目标边界框,完成异常检测一步式实现,同时该方法每秒可处理近58帧视频,满足实时性需要。本文算法在UCSD Ped1和Ped2中进行实验,并在3种不同评价准则下进行性能评估,在严格双像素准则下,Ped2中EER优于Cascade DNN 9.71%,优于Mohammad Sabokrou 13.41%,实验结果表明,本方法能够有效检测视频中异常行为且准确率较高。In order to realize the end-to-end real-time video surveillance anomalous target detection and localization in complex scenes,drawing on the target detection ideas,an end-to-end SSD real-time video surveillance anomaly detection and location algorithm is proposed.Two sets of 3×3 convolution kernels to set the target pre-selection box on the convolutional feature map of 6 different scales of the convolutional neural network are used to obtain anomaly classification and a more accurate and complete anomaly target bounding box,and the one-step implementation of anomaly detection is completed.It can process nearly 58 video frames per second,which meets the real-time requirements of anomaly detection.Different experiments are conducted in UCSD Ped1 and Ped2,and the performance is evaluated under three different evaluation criteria.Under the more strict dual-pixel criterion,EER in Ped2 is better than Cascade DNN 9.71%,and better than Mohammad Sabokrou 13.41%.Experimental results show that this method can effectively detect abnormal behaviors in video with high accuracy.

关 键 词：异常检测 目标检测 端对端 实时性 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]