在线异常事件检测的时序建模

Temporal Modeling for Online Anomaly Detection

弱监督异常事件检测是一项极富挑战性的任务,其目标是在已知正常和异常视频标签的监督下,定位出异常发生的具体时序区间。文中采用多示例排序网络来实现弱监督异常事件检测任务,该框架在视频被切分为固定数量的片段后,将一个视频抽象为一个包,每个片段相当于包中的示例,多示例学习在已知包类别的前提下训练示例分类器。由于视频有丰富的时序信息,因此重点关注监控视频在线检测的时序关系。从全局和局部角度出发,采用自注意力模块学习出每个示例的权重,通过自注意力值与示例异常得分的线性加权,来获得视频整体的异常分数,并采用均方误差损失训练自注意力模块。另外,引入LSTM和时序卷积两种方式对时序建模,其中时序卷积又分为单一类别的时序空洞卷积和融合了不同空洞率的多尺度的金字塔时序空洞卷积。实验结果显示,多尺度的时序卷积优于单一类别的时序卷积,时序卷积联合包内包外互补损失的方法在当前UCF-Crime数据集上比不包含时序模块的基线方法的AUC指标高出了3.2%。

Weakly supervised anomaly detection(WSAD)is a challenging task in that there is only normal and anomaly video label supervision but it is required to localize intervals where anomalies take place.We employ multiple instance learning(MIL)network for weakly supervised anomaly detection,which regards the input video as a bag and the segments chunked from the vi-deo as instances in it.We train the instance classifier with only label of video level(bag level),while the label of instance level is unknown.As there is strong temporal information in videos,we focus on temporal relationship for online anomaly detection in surveillance videos.We consider both global and local perspective and use self-attention module to learn each instance weight.We get the linear weighted sum of self-attention score and instance anomaly score,which represents video level anomaly score.Then the mean square error loss is employed to train the self-attention module.Online constraints allow us to use historical and current video clips only,without future frames.In order to model the temporal structure of video,we introduce LSTM and temporal con-volutional network(TCN)into WSAD problem.We explore single rate dilated temporal convolutional network,and pyramid dilated temporal convolutional network(PDTCN)which fuses multi-scale feature with different rates.Experiments show that the AUC of PDTCN with complementary inner and outer bag loss is higher than that of the baseline method without temporal mode-ling by 3.2%on UCF-Crime dataset.