KHORONOS: a camera that watches like a guard

Traditional CCTV is built around one assumption: something bad will happen, and afterwards someone will scroll through hours of footage to find it. Recording is cheap, so cameras multiply. But one operator can't really watch dozens of feeds at once, so in practice most incidents are found after the fact, if they're found at all.

KHORONOS (previously called LensGuard in our research) turns that around. Instead of a camera that records for a person to review later, it's a camera that watches the way a trained guard would: it notices people, movement and behaviour, and it speaks up while something is still happening.

Human-like attention, machine-level stamina

• It knows what normal looks like. Zones, hours and expected activity are configured per site, so a person in the stockroom at 2 a.m. is treated differently from one at 2 p.m.
• It tracks, not just detects. People and objects are followed across frames, so KHORONOS can reason about loitering, tailgating or items leaving a shelf.
• It escalates with context. Alerts carry a short clip and a plain-language description, so whoever responds knows what they're walking into.

Why the thinking happens at the edge

Streaming every camera to the cloud for analysis is expensive and slow, and it depends on a connection that isn't always there. KHORONOS runs its first layer of vision models on hardware next to the cameras. Only meaningful events, not raw video, travel upstream for deeper analysis, storage and cross-site correlation. That's how the system keeps detection-to-alert latency under 50 milliseconds on the edge and cuts bandwidth dramatically.

Privacy by design

A system that sees everything needs strict rules about what it keeps. People who aren't part of an incident are never stored as identifiable records, and incident media goes into an append-only event ledger with access controls and retention limits. That protects people's privacy and makes the evidence trustworthy.

You can read the full technical study, covering the architecture, the data design and the ROI model, in our research centre.