A cascaded U-nets model based on hybrid attention for video anomaly detection

The aim of video anomaly detection is to identify events that deviate significantly from normal behavior. The majority of the current methods for identifying anomalous behaviors rely primarily on frame reconstruction and frame prediction. However, due to the powerful representation and generation of deep neural networks, it is not sufficient to assume that anomaly events in the reconstruction methods have larger reconstruction errors. Likewise, prediction methods are not robust to noise, which can impact the effectiveness of anomaly detection. In this paper, we propose an anomaly detection method based on cascaded attention U-Nets, combining the advantages and balancing the disadvantages of the reconstruction methods and the prediction methods. Specifically, by integrating the spatial attention module and transformer-based channel attention module into the cascaded U-Nets, the weights of the spatial regions and the channel regions that represent background features can be decreased. Furthermore, to fully utilize the temporal information, a temporal shift module is incorporated into the prediction network to improve future frame prediction. Extensive experiments were conducted on several public benchmark datasets, such as the UCSD Ped dataset, the CUHK Avenue dataset, the ShanghaiTech dataset and the UCF-crime dataset. The results indicate that the proposed method can effectively enlarge the gaps between the normal events and the anomaly events, and achieve AUC 95.6% on the UCSD Ped2 dataset, 84.1% on the UCSD Ped1 dataset, 87.6% on the CUHK Avenue dataset, 75.1% on the ShanghaiTech dataset, 73.5% on the UCF-crime dataset, which outperform other previous state-of-the-art methods.