【摘要】 针对现有视频异常检测中基于预测的方法将短序列作为输入，导致序列中部分异常运动被代入到预测的下一帧，削弱了对异常运动模式的检测，文中提出一种双分支结构，每个分支专注于视频的不同维度来增强检测的准确性。第一个分支通过增强空间注意力模块和时序移位注意力模块深度挖掘视频的空间和时间特征来预测未来帧；第二个分支基于对象的光流信息挖掘视频的运动特征。最终根据时空和运动特征综合评估，以此来确定每一帧的异常情况。使用3个标准的基准数据集进行系统评估，在UCSD Ped2、CUHK Avenue和ShanghaiTech上，AUC达到99.5%、89.9%和78.6%，说明该算法在监控视频检测中具有较好的效果。更多还原

【Abstract】 In the existing prediction-based methods in the video anomaly detection, short sequences are adopted as input, which causes some abnormal motions in the sequence to be carried into the next predicted frame, weakening the detection of abnormal motion patterns. Therefore, a dual-branch structure is proposed. In the structure, each branch focuses on different dimensions of the video to enhance the accuracy of detection. The first branch is used to deeply mine the spatial and temporal features of the video by enhancing the spatial attention module and the temporal shift attention module, so as to predict future frames. The second branch is used to mine the motion features of the video according to the optical flow information of the object. Finally, a comprehensive evaluation is made based on the spatiotemporal and motion features to determine the abnormality of each frame. A systematic evaluation is carried out on three standard benchmark datasets, with AUCs of 99.5%, 89.9% and 78.6% on UCSD Ped2, CUHK Avenue and ShanghaiTech. It shows that the proposed algorithm has a good effect in surveillance video detection.更多还原