Detection of abnormal crowd behavior based on spatial-temporal adversarial variational autoencoder

Video-based detection of abnormal crowd behavior is important for the early discovery of safety risks and the prevention of group safety accidents. To address insufficient direct learning of the representation of abnormal samples because of the scarcity of abnormal crowd behavior events and the low detection accuracy of abnormal events, this study proposed a predictive spatiotemporal adversarial variational autoencoder (ST-AVAE) video anomaly detection model based on variational autoencoder, by adding the long short-term memory and adversarial network modules. Joint feature representation and reconstruction of the temporal and spatial dimensions of normal sample video sequences were performed, which reduced the feature loss in the reconstruction process of normal samples, thereby expanding the prediction loss of abnormal samples, avoiding dependence on abnormal samples, and realizing the detection of the abnormal behavior of crowd dispersal based on model reconstruction errors. Comparative experiments were conducted on the public dataset UMN and captured video datasets to prove that the ST-AVAE model has the optimal detection accuracy and recall rate in the detection of abnormal crowd escape behavior based on surveillance video, and the adversarial network module significantly improves the performance of anomaly detection.