Video anomaly detection method based on frame differential residual attention mechanism
Haiyan Cui, Bing Xu · 2023
An essential function of video surveillance systems that are widely used in fields like public safety is automatic anomaly detection. To efficiently employ spatial-temporal data, the system in this thesis combines geographical branching and temporal branching in a single network. The network has a residual self-encoder structure that, along with a DMA-based encoder and a multi-stage channel attention-based decoder, first uses a spatial-temporal wavelet analysis module for multi-frequency decomposition of video content to improve the model's understanding of high-frequency detail information and processive motion. The model learns more effectively thanks to the implementation of a channel attention mechanism that takes advantage of the channel correlations between features. Finally, training is carried out without supervision. While context dependency is retrieved from the channel attention module, the time-shifted technique is employed to take advantage of temporal aspects. The UCSD dataset is utilized to assess the system's performance. The outcomes demonstrate that our network outperforms current approaches.