ADME: Anomaly Detection in Unknown Environments via Multi-LLM Experts

Video anomaly detection (VAD) in unknown environments is a challenging task. The existing VAD methods rely on training deep models on specific datasets, which limits their generalization ability, especially when facing unknown real-world environments and different types of anomalies. To overcome this limitation, this paper presents a novel anomaly detection method via multi-LLM experts (ADME) in unknown environments. Our method leverages LLM to intelligently generate specific rules that match the current environment and achieve anomaly detection by building a framework composed of multi-LLM experts. Upon deployment in an unknown environment, we first complete the environment analysis through scene detection and infer the most probable name of the current environment. After generating specific rules, we then detect the key features of the input image. We further integrate the detected information into the chain-of-thought (CoT) to enhance the reasoning ability of LLM experts and perform anomaly detection. Experiments on the UCF-Crime and XD-Violence datasets show that our method has competitive results without any data collection and training. We deploy ADME on a robot system and verify it in real-world environments, proving its effectiveness, strong adaptability, and broad application prospects.