Video Surveillance Analytics
Defence · 2016-2019
The problem
A secured site has more cameras than any operator can watch. The overwhelming majority of the time nothing happens; the rare event that matters - a person over the fence, a vehicle at a gate it has no business being at - is buried in hours of empty footage. A wall of monitors does not scale, and constant human attention is both expensive and unreliable: vigilance fades, and the one frame that mattered goes by while someone is looking at the next screen over.
The production system watches the feeds so the operator doesn't have to. It detects objects, tracks them over time, tests them against operator-drawn restricted zones, and raises a governed alert only when something genuinely crosses a line and stays there. The hard part, exactly as with any always-on detector, is restraint: a system that alarms on every shadow and stray cat trains its operators to ignore it, so the real value is in the false alarms it doesn't raise.
The solution (downscaled)
This is a faithful, downscaled reimplementation on synthetic, iconographic perimeter frames - no real footage, no faces, no people. One vision stage reads each frame; everything else (filtering, tracking, zone tests, dwell confirmation, the alert rules, governance and dispatch) is deterministic. The product surface is the control-room video wall with the full alert lifecycle.
The honest cost story is built in: only the vision stage is metered, a change-gate skips unchanged frames, and a dwell window kills single-frame false positives before any alert reaches an operator. And there's a property that matters specifically for a secured site - in OSS mode the vision model runs on the internal network, so nothing leaves the perimeter.
What it is not
No real cameras, no live video, no real alerting integration, no real site. Every frame is synthetic and labelled; the client is not named. This is a modern reimplementation on today's stack, not a port of any legacy on-prem detector. The real detector, tracker and integrations are described in Architecture, Out of scope, and the figures are representative ranges.