Appearance-Motion Co-memory Network for Video Anomaly Detection

Abstract

Recently, video anomaly detection is currently a challenge and has attracted much attention from many researchers, which apply in the variation field like traffic accident detection, violence detection, intrusion detection systems, surveillance systems. The most common approach adopted the convolutional autoencoder that fused with appearance and motion representation to enhance the model’s ability to describe each ordinary object’s spatial and temporal behavior and quantifies the predicted error during the testing process. However, the drawback of this approach is the limit number of normal patterns which a model can learn. When training with a considerable amount of normal data, information about the normal pattern recorded in the hidden cells will be compressed, leading to missing or misleading information. This limitation is handled by a completely new improved model that applies memory modules to both the motion-appearance network and shares the same encoder, decoder. The testing on the two public datasets has shown that our model is efficient and indicates significant results improvements.