Weakly Supervised Video Anomaly Detection with Large Language Models Knowledge Enhancement Framework

Weakly supervised video anomaly detection (wVAD) aims to identify anomalous segments in videos using video-level labels. It is typically defined as a multiple instance learning (MIL) problem. However, this binary constraint method cannot learn the specific semantics of anomalies, resulting in poor interpretability of the prediction results. Additionally, the durations of different anomalous events vary, and existing anomaly event labels cannot clearly identify the contextual boundaries of anomalies. In this paper, we propose a large language models knowledge-enhanced MIL detection framework. In the proposed framework, we use large language models to obtain anomaly context boundary prompts, thus achieving clear anomaly context boundaries. By using the Anomaly Temporal Feature Amplification and Anomaly Knowledge Enhancement modules to mutually enhance each other, our framework can effectively distinguish anomalous events from non-anomalous events. Experimental results on three public benchmark datasets demonstrate the effectiveness of our method.