A Fully Connected Network Based on Memory for Video Anomaly Detection

The study of video anomaly detection (detecting abnormal events in videos) has attracted a lot of attention in the fields of computer vision and deep learning. In general, auto-encoders based on memory architecture are the mainstream anomaly detection methods. The model records the diversity of normal samples by introducing a memory module with multiple memory items. These items are used to record the different features, and participate in the reconstruction phase of the video frame. Since the reconstructed frame is mainly implemented by the convolutional layers in auto-encoder, and the Convolutional Neural Network has powerful representation capacity so that abnormal frames can also be reconstructed well by auto-encoder. By analyzing the advantages of the fully connected layers in Convolutional Neural Network, we propose an unsupervised learning method termed fully connected network based on memory for video anomaly detection. In order to reduce the representation capacity of Convolutional Neural Network, we introduce the improved Fully Connected Network that is based on the memory module. The training of the Fully Connected Network relies on the training results of the memory module, so we use a two-step scheme to train our model. Experimental results proved that our method outperforms state-of-the-art methods.