基于记忆引导双流时空编码网络的视频异常检测

Video anomaly detection based on memory-guided dual-streamspatio-temporal coding network

由于监控视频异常事件的不可知性和异常环境的复杂性,视频异常检测备受关注。当前,视频异常检测往往利用无监督的方法获取视频信息,但在特征提取过程中缺乏时空信息的获取,导致时空特征不一致的问题。为此,提出一种基于记忆引导的双流时空编码器网络(MSTAE)模型,设计了一种双流时空特征提取网络,分别以连续的视频帧序列和光流图为输入,空间流获取视频的运动特征,时间流获取视频的时序特征,同时,引入注意力机制改进编码器,降低因数据冗余导致的风险误差。在3个公开标准数据集(Ped2、Avenue和ShanghaiTech数据集)上进行了广泛的实验,结果表明,模型的AUC精度优于目前大多数的方法。

Video anomaly detection has attracted much attention due to the unpredictability of surveillance video anomaly events and the complexity of the anomaly environment.Currently,video anomaly detection often utilises unsupervised methods to acquire video information,but the lack of spatio-temporal information acquisition during feature extraction leads to the problem of inconsistent spatio-temporal features.To this end,a memory-guided two-stream spatio-temporal coding network(MSTAE)model based on memory-guidance is proposed,and a two-stream spatio-temporal feature extraction network is designed to obtain the motion features of the video in spatial streams and the temporal features in temporal streams using continuous video frame sequences and optical flow graphs as inputs,respectively,and at the same time,the attention mechanism is introduced to improve the encoder and reduce the risky error due to the data redundancy.Extensive experiments are conducted on three open standard datasets(Ped2,Avenue and ShanghaiTech datasets),and the experimental results show that the AUC accuracy of the model outperforms the state-of-the-7art methods.