Pseudo-anomaly-guided convolutional autoencoder video anomaly detection

Video anomaly detection plays an important role in national security and medical monitoring because it can maintain public safety efficiently and at low cost. Reconstruction-based deep autoencoder anomaly detection methods have been extensively studied for their powerful representational capabilities. However, autoencoder can often also successfully reconstruct abnormal behavior, leading to missed detection of abnormal behavior. To solve this problem, a pseudo-anomaly guided convolutional autoencoder video anomaly detection method is proposed. The model uses 3D convolution to extract video spatio-temporal features. Firstly, the normal data is used to simulate the abnormal data distribution to generate pseudo-abnormal data. Two methods of generating pseudo-anomalies are proposed: skip frame and patch based. Then, the normal data and the generated pseudo-abnormal data were used to train the model. During the training, the normal data was reconstructed well while the pseudo-abnormal data was reconstructed poorly. Therefore, the model was encouraged to limit the reconstruction of abnormal data. Finally, compared with other reconstruction-based models on UCSD-Ped2, Avenue and ShanghaiTech public video anomaly detection data sets, the detection accuracy has been effectively improved.