False-Positive Taxonomy for Safety AI

Who this is for

For safety and operations teams reviewing nuisance alerts during pilots.

The buyer question

How should false positives be categorized before tuning a safety AI system?

Scene causes

Lighting, glare, weather, camera angle, occlusion, dust, steam, and reflective surfaces should be separated from model or workflow issues. A forklift charging bay that steams every shift change will generate the same nuisance pattern every day until it is tagged as a scene cause, not a model failure.

Rule causes

Some alerts are technically correct but operationally wrong because the zone, schedule, allowed behavior, or threshold was defined poorly. A maintenance crew working inside a mapped restricted zone during an approved window is a rule cause, not a detection failure, and the fix is a schedule exception, not a model retrain.

Workflow causes

An alert can feel false if it reaches the wrong person, uses the wrong priority, or lacks location context.

How to use this with DHI

Use this page as a pre-pilot checklist. Pick one zone, one event type, one alert owner, and one review cadence. If the current cameras cannot support the workflow, fix the camera plan before expanding the deployment.