What is AI video analytics, and what can it reliably do?
What is AI video analytics?
AI video analytics is software that analyzes camera video to flag a short list of defined events, such as a person entering a restricted zone or a blocked exit, and sends each one to a person to review. It does not understand scenes as people do; it finds only what it was set up and tested to find.
From stream to alert
- Read the streamThe software reads video from an IP camera or recorder, usually over a standard streaming protocol such as RTSP, on a server on site or in the cloud.
- Find objectsA trained model looks for the object types it knows, such as people, vehicles, hard hats or pallets, and marks where they are in each frame.
- Apply your rulesEach detection is checked against zones drawn on the image, schedules and minimum durations, so a brief pass-through is not an event.
- Send a clipWhen a rule is met, a short clip with the time, camera and a confidence level goes to a person.
- Learn from reviewsConfirmed and dismissed alerts show where a camera or rule needs adjusting. On well-run systems this is done by people, deliberately, not silently.
Common misunderstandings
- Not general understandingIt does not know what is happening in a scene. It matches patterns it was trained on, inside rules you wrote.
- Not face recognition by defaultMost safety and operations uses need no identity at all. Face recognition is a separate, high-risk capability with its own legal rules.
- Not a replacement for CCTVIt runs on top of cameras and recorders. The footage remains the record.
- Not proof that all is wellNo alert does not mean nothing happened. A camera can be blocked, dark or offline, and a good system says so.
What can AI video analytics be used for, and what still needs a person?
Examples describe the category in general, not any product. Each use needs testing on the cameras and scenes where it will run.
Why the same rule works on one camera and not another
- Angle and distanceObjects that are small, far away or seen from directly above or below are hard to detect.
- LightNight, glare and backlighting reduce what can be seen. Many cameras switch to black and white at night, which removes color cues.
- OcclusionPeople and objects hidden behind racking, vehicles or other people cannot be detected.
- Scene changesMoved shelving, new signage or a repainted floor can change results after set-up.
- Stream qualityHeavy compression, low frame rates and dropped streams all lower what the software sees.
Questions that separate a demo from a deployment
- Results on our footageAsk for misses and false alerts measured on your own cameras, with the method, sample and date stated. A general accuracy number tells you little.
- Camera by cameraAsk how each camera will be tested, and what happens to cameras that fail.
- Who reviewsAsk who sees each alert, how repeats are throttled, and how dismissals are recorded.
- SilenceAsk how the system shows a camera that is offline or blocked, so silence is not read as safety.
- DataAsk what video is kept, for how long, where, and who can see it.
What to settle first
- PurposeWrite down which events you will analyze and why, and keep the list short.
- NoticeTell people in the area that cameras are in use and, where required, that video is analyzed.
- AssessmentIn many places a documented risk assessment is expected before systematic monitoring. Take qualified advice for your market.
- WorkersConsult employees or their representatives where required, and keep analysis about places and events, not individuals.
How do we test AI video analytics on our site?
- 1Pick two or three eventsThe events you most want to hear about the same day, each with a zone and hours.
- 2Test each cameraCheck angle, light and stream quality for each event, and drop cameras that fail.
- 3Measure on your footageRun on your recordings first and write down misses and false alerts with a stated method.
- 4Run with a named reviewerGo live only where results are acceptable, with every alert reviewed by a person.
Last updated
Questions, answered.
CCTV records video for people to watch or search. AI video analytics runs on the same video and flags defined events for a person to review. Most sites use both.
Often, but not always. Each camera needs a suitable view and enough light for the event you care about, so test cameras one by one before relying on them.
There is no honest general figure. Results depend on the camera, scene, light and event. Ask for results measured on your own footage, with the method stated.
It does not have to. Zone entry, blocked exits, queues and PPE checks need no identity. Identification is a separate capability with stricter legal limits.
Agentra’s AI Vision Safety, which reviews defined safety events on tested cameras, is planned and not available to buy yet.
Bring us one process. We’ll show you how it would run.
A 30-minute call with a specialist: what it connects to, what a pilot involves, and an honest view on fit, including when it isn’t one.
- 30 minutesOne call, one process
- No commitmentNo sales script
- 4 business hoursTypical reply time · Mon to Fri, 7:00 to 22:00 IST
- 1You get a short agenda by email.
- 2We look at one process with you and how it would run.
- 3You get an honest view on fit, and what a pilot would involve.