Video Anomaly Detection

Video anomaly detection aims to find and sometimes localize events that differ from expected behavior in settings such as surveillance, factories, and driving. Common approaches learn normal patterns from reconstruction or prediction, model motion with poses or trajectories, or use weak labels and multiple-instance learning; newer systems also apply vision-language models for zero-shot reasoning and explanations. UCF-Crime, ShanghaiTech, and XD-Violence are widely used benchmarks, but results often depend on frame-level AUC, which may not reflect accurate event localization. A central challenge is making systems generalize to new scenes and unfamiliar anomalies while improving temporal grounding, recall, and reliable evaluation.