基于双支路的人群视频异常检测算法研究

Research on Crowd Video Anomaly Detection Algorithm Based on Dual-branch

该文对人群异常检测任务进行了研究,针对人群场景视频背景较冗余,易受光线与噪声影响以及实际部署等问题,提出了人群异常多尺度特征记忆网络(CaMsm-net)。为了从多角度判别异常的发生,并将各类信息更好地融合,该网络采用了一种双支路共享单元结构,将原始帧与经背景建模后的帧同时输入网络结构中,分别对两条支路进行预测,采用预测误差进行异常的判定,并且从实际应用角度出发,框架结合了深度可分离卷积、数据扩增等方法,保障了检测的准确性与部署的可行性。在公开的美国明尼苏达大学(UMN)人群数据集与实际监控下的火车站出站口数据集中的实验表明,曲线下面积(AUC)指标分别达到了99.2%与84.1%,平均检测准确率分别为95.9%与81.7%,证明了所提算法能够较好地检测到各类人群异常现象的发生,并具有更广的适用性。

This paper studies the task of crowd anomaly detection.Considering the problems of crowd scene video background being redundant,susceptible to light and noise,and actual deployment,a Crowd anomaly Multi-scale feature memory network(CaMsm-net)is proposed.In order to distinguish the occurrence of anomalies from multiple angles and integrate better various types of information,a dual-branch shared unit structure is adopted by the network,the original frame and the background modeled frame are simultaneously input into the network structure.To predict the two branches separately,the prediction error was used to determine the abnormality,and from the perspective of practical application,the depthwise separable convolutionand data augmentation methods are added to the framework to ensure the accuracy of detection and the feasibility of deployment.Experiments on the public University of MinNesota(UMN)population dataset and the actual monitoring of the train station exit dataset show that the Area Under Curve(AUC)indicators reach 99.2%and 84.1%respectively,and the average detection accuracy rates are 95.9%and 81.7%,which proves the proposed algorithm can better detect the occurrence of various crowd abnormalities and has wider applicability.