Summary
Video anomaly detection finds unusual events in video footage, but most methods need lots of training and labeled examples, which limits their use in new situations. The authors present Cog-VADU, which uses large vision-language models to reason through video segments step-by-step, like a chain of thoughts, without any training. This approach helps the system remember what happened before, making it better at spotting complex anomalies versus normal fast actions. They also add a stage that checks consistency between text explanations and video data to make the detection more reliable. Their tests show competitive results on popular benchmarks and that the reasoning approach helps across different models.
What this means in practice
- •For security monitoring teams: Detect unusual activities in surveillance footage without needing tailored training for each site, improving adaptability and interpretability.
- •For video analytics developers: Integrate a training-free reasoning module that improves anomaly detection consistency and explanation across different vision-language models.
Authors
Mohd Ubaid Wani, Sara Atito, Josef Kittler, Muhammad Awais
Abstract
Video Anomaly Detection (VAD) aims to temporally localize abnormal events in videos. Most existing approaches rely on dataset-specific training and curated annotations, limiting generalization in open-set scenarios. Recent zero-shot methods based on Large Vision- Language Models (LVLMs) alleviate this dependency but often lack temporal continuity and structured reasoning. We propose Cog-VADU, a fully training-free framework that reformulates VAD as a sequential cognitive reasoning task. Cog-VADU introduces Chain-of- Anomaly Detection Thought Prompting (CoADTP), which unrolls an LVLM into a recurrent reasoning chain across video segments. By propagating structured rationales over time, the model maintains implicit temporal memory, enabling robust discrimination between com- plex anomalies and high-motion normal activities. To improve reliability, we further design a cross-modal re-ranking stage that aligns textual rationales with visual embeddings, enforcing semantic consistency and temporal coherence for refined and stable predictions. Extensive experiments on multiple public VAD benchmarks demonstrate that Cog-VADU achieves competitive zero-shot performance. Moreover, cross-model evaluations show that CoADTP consistently enhances reasoning-based anomaly detection in a model-agnostic manner, pro- viding interpretable and generalizable anomaly understanding for real-world applications.