Global Temporal Attention-Driven Transformer Model for Video Anomaly Detection

Video anomaly detection is an important task in the field of computer vision and is widely used in scenarios such as intelligent monitoring, security prevention, and behavior analysis. Traditional methods have limitations in dealing with long-term dependencies and modeling global temporal information, making it difficult to accurately identify complex abnormal behaviors. To this end, this paper proposes a video anomaly detection method based on Transformer and global temporal attention mechanism to improve the modeling ability of long-term dependencies. Specifically, the Transformer structure is first used to capture the global relationship between frames, and the global temporal attention mechanism is introduced to optimize the extraction of key time step information. The experiment is conducted on the public dataset and compared with methods such as Conv2D-AE, STAE, ConvLSTM-AE, TSC, MemAE, MNAD-AE, and Stack-RNN. The results show that Ours method achieves the best performance in terms of AUC and EER indicators. In addition, the training loss curve and anomaly score visualization analysis further verify the stability and effectiveness of the model. Future research can explore more efficient attention mechanisms and multimodal fusion to further improve anomaly detection capabilities.