Temporal-spatial interactive shift module for videos anomaly detection

Video anomaly detection attempts to identify abnormal activity within an overwhelming volume of surveillance videos. In order to learn efficient features in both spatial and temporal dimensions for anomaly detection with a lower calculation cost, a temporal-spatial interactive shift module (TISM) is proposed, avoiding the huge computation cost of 3D convolution. The module shifts part of the feature channels over the time dimension via weighted spatial interaction with temporally neighbored features, allowing spatial information to be interactively exchanged between adjacent frames. Entropy is introduced into the interactive operation for channel selection and adaptive weighting. Experiments on the UCF-Crime dataset prove the superiority of the proposed method. The AUC is increased by 2.53% compared with the SOTA baseline and 6.24% compared with the 3D CNNs, while reducing the amount of calculation by 26%.