Video anomaly detection network based on LWSMHL dual memory branching

In order to solve the defects of video anomaly detection based on predictive frame method which is weak in dynamic and spatial information extraction, a dual memory branch video anomaly detection model (light weight spatiotemporal memory-high level memory, LWSM-HL) is designed. The designed model considers both spatio-temporal features and global background features. The LWSM-HL model consists of two branches: A branch is based on an improved long-term and short-term memory network that repeatedly loops to extract spatiotemporal information; another branch extracts global background features to match those in the externally high-level memory. The two branch features are fused and then the video frames are predicted for output by the encoder. Effective global high-level background features are memorized in high-level memory for more scenarios. Experimental results on public datasets show that the designed video anomaly detection model has some advantages, reaching an AUC of 87.3% on the CUHK Avenue dataset. The designed model not only enhances the representation of spatial motion features, but also attaches global background information to the predicted frames, thus effectively improving the accuracy and robustness of video anomaly detection.