A Novel Spatio-Temporal Information Knowledge Distillation Strategy for Weak-Supervised Video Anomaly Detection

Video anomaly detection (VAD) is highly valuable for applications in intelligent monitoring systems. However, due to the complexity of video, as well as the lack of detailed annotation information, VAD remains a challenging problem. Although existing self-training-based methods make process in anomaly localization and reduce the need for annotation information, they still suffer from two limitations: 1) insufficient spatio-temporal information correlation in two-stage model; 2) ignoring label noise. To address the above issues, we propose a novel self-training video anomaly detection framework based on spatio-temporal information knowledge distillation (STKD-SVAD). Specifically, we design a spatio-temporal information knowledge distillation strategy to establish correlations for spatial and temporal information. Furthermore, we design a temporal information module (TIM) and a spatial information module (SIM) to enhance the temporal and spatial information, respectively. Moreover, we add multi-instance loss in the second-stage of self-training framework to reduce the interference of label noise. Experimental results on public datasets demonstrate the effectiveness of our method.