Dual-correlation graph attention network for video anomaly detection

Anomalous events in videos often exhibit strong spatial-temporal correlations. However, existing video anomaly detection methods still struggle to model complex interactions between different regions within and across frames. To address this challenge, we propose a dual-branch spatial-temporal graph convolutional network (STGCN), which effectively captures both spatial structure and temporal evolution in video sequences. Input frames are divided into multiple patches and encoded into local feature units. These feature units are then passed into two parallel graph branches. The inter-correlation branch constructs a spatial graph based on feature similarity across patches, modeling interactions between different spatial regions, while the intra-correlation branch builds a temporal graph based on sequential continuity, capturing the evolution of local features across consecutive frames. To integrate spatial and temporal information more effectively, a learnable weighting mechanism is introduced to fuse the outputs of the two branches, generating enhanced representations. These representations are decoded to predict future frames. Experimental results on two public anomaly detection datasets (ShanghaiTech and UCSD-Ped2) demonstrate that the proposed method achieves superior performance compared with the state-of-the-art approaches.