Video Sequence Anomaly Detection with Multi-Layer Memory-Augmented Autoencoder
Minxiang Long · Academic Journal of Computing & Information Science · 2023
In order to enhance the model's ability to learn normal pattern features in the video anomaly detection task, this paper proposes an end-to-end video anomaly detection method that combines reconstruction and prediction. The method consists of two modules: (1) multi-layer memory-enhanced auto-encoder module: reconstructs RGB frames using a multi-layer auto-encoder with skip connection to compensate for the information loss due to memory; (2) conditional variational auto-encoder module: the reconstructed RGB frames from the previous step are taken as inputs, and predicts future frames using the current optical flow as a condition to capture the correlation between the optical flow and the video frames. Comparative experiments are conducted on Avenue, Ped2, and SHTech datasets, and the experimental results show that the hybrid model achieves relatively strong anomaly detection capability.