Improving Anomaly Detection in Videos: An End-to-End Local-Global Temporal Dependency Network

As both video sequences and anomalous events are composed of continuous segments, there exists a temporal dependency between these segments. Therefore, simultaneously learning the local temporal dependency between adjacent segments within a video and the global temporal dependency between adjacent segments can improve the detection performance of the model. However, previous weakly supervised methods have ignored the long-term and short-term temporal dependencies between segments. To alleviate this issue, this paper proposes an end-to-end local-global temporal dependency network by introducing temporal dependencies between segments, enabling the model to focus more on the correlations between segments.