LTGS-Net: Local Temporal and Global Spatial Network for Weakly Supervised Video Anomaly Detection

Video anomaly detection has an important application value in the field of intelligent surveillance; however, due to the problems of sparse anomaly events and expensive labeling, it has made weakly supervised methods a research hotspot. Most of the current methods still adopt the strategy of processing temporal and spatial features independently, which makes it difficult to fully capture their temporal and spatial complex dependencies, affecting the accuracy and robustness of detection. Existing studies predominantly process temporal and spatial information independently, which limits the ability to effectively capture their interdependencies. To address this, we propose the Local Temporal and Global Spatial Network (LTGS) for weakly supervised video anomaly detection. The LTGS architecture incorporates a clip-level temporal feature relation module and a video-level spatial feature module, which collaboratively enhance discriminative representations. Through joint training of these modules, we develop a feature encoder specifically tailored for video anomaly detection. To further refine clip-level annotations and better align them with actual events, we employ a dynamic label updating strategy. These updated labels are utilized to optimize the model and enhance its robustness. Extensive experiments on two widely used public datasets, ShanghaiTech and UCF-Crime, validate the effectiveness of the proposed LTGS method. Experimental results demonstrate that the LTGS achieves an AUC of 96.69% on the ShanghaiTech dataset and 82.33% on the UCF dataset, outperforming various state-of-the-art algorithms in anomaly detection tasks.