An Attention-based U-Net Network for Anomaly Detection in Crowded Scenes

Abstract

Anomaly detection in surveillance video is of great significance for public safety. Deep autoencoder has been widely used in anomaly detection. Because of its good generalization ability, sometimes abnormal samples can still be reconstructed very well. Some scholars use memory module constructed by using the normal samples to reconstruct the test samples. The memory items need to be retrieved and updated during the training and testing process, hence more memory space is required to store memory module, which greatly increases the training and test costs. We tackle the problem of the excessive generalization ability of autoencoder from a new perspective. We introduce an attention mechanism to propose an attention-based U-Net network to detect anomalies. The network adds an attention module before the skip connection of U -Net network, so that the model pays more attention to the foreground targets. During the training process, the normal foreground targets are learned more fully. Therefore, in the test, the proposed method can achieve more accurate prediction of normal targets that appear frequently, so that the rare abnormal targets can be highlighted because of the large prediction errors. We conduct experiments on real surveillance videos of UCSD Ped l and ShanghaiTech datasets, and the experimental results demonstrate that the proposed method is an efficient model.