Aegis Vision

Transparency

What Aegis Vision measures, how it measures it, and what it will publish about the detection system as deployments are evaluated.

A detection system that schools rely on should be described the way a safety-critical product is described: what it is for, what it is not for, how it is tested, and what the tests showed. This page sets out that structure. Figures are published here as deployments are evaluated, each with its test conditions and date.

Intended use

Real-time AI detection of weapons and safety incidents on a site's existing cameras, processed on the site's own hardware. It is designed for Schools and universities, operated by the school's own security staff, with every alert reviewed by a person before any external escalation.

Out of scope

Airports; identifying individuals. The system is not a law-enforcement tool and is not used to monitor individuals. Weapon and incident detection works on objects and body movement. It does not identify people, match faces against any database, or build profiles.

System description

A multi-model computer-vision system of more than one billion parameters in total. Detect, verify, alert, respond. A fast detector scans every frame; a larger model verifies the detection and the alert is raised to the site's guard console; two classifier models then grade the incident; a guard confirms, dismisses or escalates. Three independent checks before an alarm sounds; borderline detections must persist across frames and be confirmed by a second model, and the classifier stage can stand an alarm down while keeping the event logged.

How performance is measured

  • Held-out footage. Evaluation uses recordings from cameras that were not used to tune the model, drawn from the deployment being evaluated.
  • Per class. Precision and recall are reported separately for each weapon class and each incident type, with confidence intervals and the number of test samples.
  • False alarms per camera per day. Counted on continuous footage, not on curated clips, because this is the figure that decides whether guards keep trusting the system.
  • Time to alert. From the first frame in which the object or event is visible to the alert appearing on the guard console, reported as aggregate statistics across deployments and never at a level that reveals a single site's response window.
  • Behaviour modules. Behaviour modules have no published accuracy figures yet; figures will be published once measured on real deployments.

What is published, and when

Each published evaluation carries the software and model version, the date, the camera and scene conditions in general terms, and the sample sizes. Nothing is published that would help someone evade detection: no detection limits by object, angle, lighting or distance, no site layouts, no response-time windows for individual sites.

Results will appear on this page as evaluations complete. Questions about methodology: help@aegisvision.org.

Updated 2026-11-17