LLMs enhance video anomaly detection with reasoning and spatial grounding

Researchers have developed VANGUARD, a novel framework that integrates video anomaly detection with multimodal large language models. This system not only identifies anomalies but also provides interpretable chain-of-thought reasoning and precise spatial localization of the anomalous events. VANGUARD utilizes a staged training approach and a teacher-student annotation pipeline, achieving strong performance on benchmarks like UCF-Crime and demonstrating cross-domain generalization. AI

影响 Introduces a new method for interpretable video anomaly detection, potentially improving surveillance and security systems.

排序理由 This is a research paper detailing a new framework for video anomaly detection using multimodal large language models. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.CV 阅读 →

• GroundingDINO
• Qwen3-VL-4B
• ShanghaiTech
• UCF-Crime
• VANGUARD
• XD-Violence

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →