融合记忆增强的视频异常检测

深度自编码器可以通过预测当前帧来判断视频中的异常情况。但由于自动编码器对图片的低层次特征无法良好的预测，在自动编码器中添加跳跃连接可以提高预测图片细节信息的能力。由于自动编码器有很好的“泛化”能力，为了抑制对异常事件的准确预测，通过在编码器和解码器的跳跃连接之间添加记忆增强模块限制模型对异常帧的准确预测。同时，为了突出异常帧中的事件，在预测视频帧的同时通过背景提取模块获取当前图片的背景信息用于后续预测误差的计算。在UCSD Ped2数据集、CUHK Avenue数据集和ShanghaiTech数据集上的实验结果表明，改进后模型的异常检测能力得到了提升。

Video Anomaly Detection Combining Memory-Augmented

Deep autoencoder can predict the current frame to find abnormal events in videos. However, the autoencoder can not predict the low-level features of the image well, adding short-cut in the autoencoder can improve the ability to predict the details of the image. Autoencoder has good generalization ability, in order to suppress the accurate prediction of abnormal events, a memory augmented module is added between the short-cut of encoder and decoder to limit the accurate prediction of abnormal frames. At the same time, in order to highlight the events in the abnormal frame, the background of the current frame is obtained by the background extraction module while the video frame is predicted, which is used for the calculation of prediction error. The experimental results on UCSD Ped2 dataset, CUHK Avenue dataset and ShanghaiTech dataset show that the ability of anomaly detection is improved.