Classic video analytics only recognise the handful of events they were configured for. Describing alerts in plain language changes what a camera can watch for.
The limits of rules
Most deployed video analytics are built around a fixed menu: line crossing, loitering, a colour, a plate. Each rule is useful on its own, but the real world rarely lines up neatly with the menu.
When an operator needs to watch for something new, the usual answer is a ticket to an integrator and a wait measured in weeks.
Describing instead of configuring
Language models paired with vision let operators simply describe what they care about. The system interprets the description and applies it to live feeds.
- “A person propping open the fire exit”
- “A forklift entering the pedestrian walkway”
- “Two people exchanging a package near the gate”
What changes for operators
Alerts become something any team member can write and refine. Coverage expands to match real concerns rather than the limitations of a configuration screen.