2025
Грибова В. В., Филаретов В. Ф., Xing M., Feng Z., Su Y., Zhang Y., Oh Ch., Huang D.
Neurocomputing, Q1
Xing M., Feng Z., Su Y., Zhang Y., Oh Ch., Gribova V., Filaretoy V.F., Huang D. Spatio-temporal graph-based self-labeling for video anomaly detection // Neurocomputing. 2025. Vol. 627. 129576. https://doi.org/10.1016/j.neucom.2025.129576
Vol. 627. 129576
Статьи в журналах
Video anomaly detection (VAD) aims to identify abnormal events in a video sequence. Existing methods achieve VAD by learning the decision boundary between the normal space and the abnormal space pre-defined in the training data. However, these methods trend to neglect the distribution gap between the pre-defined abnormal space and the real one, which lead to overfitting on the normal space or bias toward the pre-defined abnormal space. In this paper, we propose a spatio-temporal graph-based self-labeling method that not only focuses on the pre-defined abnormal space but considers the real abnormal space, enabling it to capture the decision boundary between the normal space and a complementary space, called as the not-normal space. We first construct a spatio-temporal graph (ST-Graph) based on the objects of input video and utilize a spatio-temporal graph convolution network (ST-GCN) to model the interaction between objects. We then propose a self-labeling-based learning mechanism that encourages the proposed ST-GCN to record the normal events while abstaining from labeling the pseudo-abnormal events, thereby aggregating the pre-defined and real abnormal spaces into not-normal space. To evaluate the model performance on localizing anomalous objects and capturing interactions between objects, we further introduce an object-level criterion that bridges frame-level and pixel-level criteria. Our method is validated on three datasets and achieves state-of-the-art frame-level AUC results on Avenue (92.5%), and outperforms existing ST-Graph-based methods on UCSD Ped2 (96.5%) and ShanghaiTech (76.8%).