Multi-scale Siamese prediction network for video anomaly detection

Automatically detecting anomalous events in surveillance videos is crucial for security maintenance. Due to the challenging nature of the task, the performance of the existing approaches is still limited. In this study, we propose a video anomaly detection method called multi-scale Siamese prediction framework (MSSP), where the Siamese network uses the information embedded in the observed anomalous events without requiring any additional parameters. To extract spatiotemporal features, we introduce a multi-scale term where an improved inception module and a convolutional GRU (Conv-GRU) module are combined. They are employed in each layer of the U-Net coding stage to mitigate the information loss caused by subsampling. To further optimize the proposed model, a loss function combining the prediction loss function and the contrastive loss is proposed. We evaluate the system performance on three public datasets: CUHK Avenue, UCSD Ped2, and ShanghaiTech dataset. Experimental results demonstrated that the MSSP framework achieved AUC values of 89.4%, 97.4% and 73.83%, respectively, which significantly outperforms other methods.