基于双流结构的异常行为检测模型

为更好利用输入视频的时域特征,提升异常行为检测精度,采用三维自编码器为主体的网络分支编解码视频的时空域信息,提出改进光流融合策略的时域分支提供额外时域信息.将双分支结果融合并计算重建误差,在此基础上进行异常行为的判断.针对目前像素评价指标的不足,提出一种改进的像素级别检测指标.结果表明,融合后的结果好于各分支单独的结果...

ABNORMAL BEHAVIOR DETECTION MODEL BASED ON TWO STREAM STRUCTURE

In order to make better use of the time-domain information of video in abnormal behavior detection and improve the accuracy of detection,3D auto-encoder was adopted as the main part of the network branch to encode and decode the spatial-temporal information of input video block.Simultaneously,the time-domain branch with improved optical flow fusion strategy was proposed to provide additional temporal information.Then the results of two branches were fused to obtain reconstruction error to judge abnormal behavior.In addition,aiming at the defect of current pixel evaluation indication,a new pixel level detection indication was proposed.It shows that the results of fusion are better than the results of individual branch,and better than the methods in recent years.It can be concluded that the network branch and the temporal branch complements each other,further improving the overall detection effect of the model.