Video anomaly detection (VAD) systems are designed to flag deviations from normal behavior, but not every deviation corresponds to a safety-critical event. A new preprint on arXiv investigates this gap specifically for human-centered emergencies and violence.

The authors ask whether current evaluation methods are inadvertently measuring observation-quality shortcuts—such as camera angle, resolution, or occlusion—rather than the semantic content of the anomaly. This is a known risk in computer vision benchmarks, where models can exploit spurious cues.

The abstract is truncated, so the findings are not yet available. But the research direction is important: if VAD evaluations are skewed by observation quality, then reported performance may overstate real-world safety capabilities.