O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning

Industrial Video Anomaly Detection (IVAD) aims to identifyanomalous objects and events in an industrial process, which is crucial formodern manufacturing and quality control systems. Existing VLM-basedanomaly reasoning methods are capable of detecting open-ended anoma-lies in general domains. However, their performance declines in industrialsettings characterized by intricate object transformations, strict physics,and procedural constraints. To tackle the complexity of such interaction-intensive detection, we introduce a training-free agentic framework O-VAD for anomaly detection free of domain-specific knowledge, empha-sizing object state evolution like humans inspectors. It is designed totrack spatial-temporal dynamics and underlying transformations of de-tected objects over time, and then reason over the object-wise temporalstate trajectories to identify abnormal objects in grounded frames. Ourmethod overcomes limitations of prior approaches that rely on retrainingon normal clips or injecting domain knowledge as context for test-timeinference. Extensive experiments on three IVAD datasets demonstratethat our method outperforms frontier VLMs, agentic frameworks, andtraditional VAD methods fine-tuned on the respective datasets, whileproviding interpretable reports over anomaly processes and types.