摘要
在视频异常行为检测过程中,为了提取出可分辨性更好的特征,同时兼顾运行速度,提出一种基于优化的全卷积网络(full convolution network,FCN)的异常行为检测与定位方法。对FCN进行优化,使用卷积神经网络(convolution neural network,CNN)的数个初始卷积层和一个额外卷积层,生成同时描述运动和形状的区域向量集;使用高斯分类器对特征向量集进行验证,将存在显著差异的分块标记为异常,将低拟合置信度的可疑区域输入到稀疏自动编码器中;对异常行为进行定位,并将异常行为的位置传回FCN。所提方法在UCSD和Subway这2个公开数据集上进行验证分析。实验结果表明,所提方法在受试者操作特征(receiver operation characteristic,ROC)曲线、等错误率(equal error rate,EER)和曲线下面积(area under curve,AUC)性能方面表现优秀,且运行速度达到60 frame/s,实时性较为优秀。
In the process of video anomaly detection,in order to extract better distinguishable features and take into account the running speed,this paper proposes an anomaly detection and location method based on optimized full convolution network(FCN).Firstly,the FCN is optimized by using several initial convolution layers and an additional convolution layer of convolution neural network(CNN)to generate a set of regional vectors describing both motion and shape.Then,the set of eigenvectors is validated by using Gauss classifier,and the blocks with significant differences are marked as anomalies,and the suspicious regions with low fitting confidence are input to the sparse automatic encoder.Finally,the abnormal behavior is located and the location of the abnormal behavior is returned to FCN.The proposed method is validated and analyzed on two open data sets,UCSD and Subway.The experimental results show that the proposed method performs well in receiver operation characteristic(ROC)curve,equal error rate(EER)and area under curve(AUC).In addition,the proposed method has achieved a processing speed of 60 fps,indicating an excellent real time capability.
作者
陈纪铭
陈利平
CHEN Jiming;CHEN Liping(School of Computer and Information Science,Hunan Institute of Technology,Hengyang 421002,P.R.China)
出处
《重庆邮电大学学报（自然科学版）》
CSCD
北大核心
2021年第1期126-134,共9页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
湖南省自然科学基金(13JJ9027)
湖南省教育科学“十三五”规划课题(XJK18CXX013)阶段性成果
湖南工学院科学研究项目(2018HY016)。
关键词
异常行为检测
全卷积网络
定位
高斯分类器
稀疏自动编码器
anomaly detection
full convolution network
location
gauss classifier
sparse automatic encoder
