Video analytics: what it is, how it works and what it gives a business in 2026
Author: DuoVision team
For: owners and managers of enterprises, security, HACCP and occupational safety teams, and IT directors who are looking into video analytics for the first time
Summary
- Video analytics is software that watches the camera feed in place of a person and reports events you described in advance: a person without a hair net in a production area, a box on the floor, a queue at a counter, an outsider on the site.
- It differs from a motion detector in that it understands who or what is in the frame and where. So a shadow or a bird should not trigger an alarm.
- Frames can be processed in the camera, in the recorder, on a server at the site, or in the cloud. Each option has its own cost, internet requirements, and a question about where the video goes.
- Accuracy in a demo and on your site differ. In 2025 VARUS tested solutions from six companies, and only one of them gave stable video analytics.
- From 28.01.2026 Macroscop and Axxon products (Axxon Next, "Intelekt") are on the list of prohibited software. The ban is mandatory for state systems and critical information infrastructure; private businesses should plan a replacement. Cameras do not have to be replaced: the analytics takes video from the cameras you already have.
For: owners and managers of enterprises, security, HACCP and occupational safety teams, and IT directors who are looking into video analytics for the first time
What is video analytics in plain words
At most enterprises the cameras record video that nobody watches. The recording is opened when something has already happened: goods went missing, someone was injured, a complaint arrived. An operator at a console physically cannot follow twenty screens for a whole shift.
Video analytics takes over this viewing. The software receives the feed from the cameras, looks for people, objects or vehicles in every frame, and checks them against the rules you set. When a rule is broken, a person receives an event with a frame, the camera name and the time. The rest of the time the system stays silent.
A rule is worded in the language of your process: "a person without a mask in the deboning area", "a box in the aisle between the racks", "more than five people at the counter", "someone by the fence after 22:00". Video analytics only notices such situations. What to do about them is up to you.
How video analytics differs from motion detection and ordinary video surveillance
These three are often confused, because all of them work with the same cameras. The difference is in what the system understands about the frame.
| Parameter | Video surveillance | Motion detector | Video analytics |
|---|---|---|---|
| What it does | Records and shows video | Reacts to a change of pixels in the frame | Finds people, objects and vehicles and checks a rule in a zone |
| Who watches | An operator or nobody | An operator after an alarm | A person receives only the events that match the rules |
| Typical error | An event is missed because nobody was watching | An alarm from a shadow, rain, a bird, a light switched on | There are errors, but they can be measured and reduced by tuning |
| What it gives | Evidence after the event | A signal that "something moved" | A signal of "what exactly happened and where", plus an event log |
A motion detector does not know what moved. So in production, where people and machinery move all the time, it either stays silent or raises alarms without end. Video analytics can tell a person from a pallet truck and see whether the person is wearing gloves.
How video analytics works: from a frame to a person's decision
Different systems differ a lot under the hood, but the path of an event is roughly the same.
- Frame. The system takes the video stream from a camera or recorder, usually over the RTSP protocol.
- Recognition. The system finds in the frame what the rule needs: a person, protective equipment, a box, a vehicle.
- Rule in a zone. Zones are drawn on the camera frame: a workstation, an aisle, a ramp, a perimeter. The rule fires only when the condition is met right there and holds for a set time. A person who walked past a table without a hair net on the way to the changing room and a person who stands at the table without a hair net are different cases for the rule.
- Event or card. Neighbouring frames of one situation are merged into one event. A person receives a card with a frame as evidence, the camera, the rule and the time.
- A person's decision. The responsible person confirms the card or rejects it. What is confirmed goes into a log that you can work through with the team and show to an auditor.
We consider the last step mandatory. The system can make mistakes, so penalising, blocking or blaming anyone without a person looking at the frame is not acceptable.
Where video analytics runs: in the camera, the recorder, on a server or in the cloud
Where frames are processed determines the cost, the internet requirements, and whether the video leaves the enterprise.
| Where | Pros | Cons |
|---|---|---|
| In the camera | No separate server needed, video goes nowhere | Little computing power, the set of functions is defined by the camera maker, usually it cannot be trained for your task, each camera has its own logic |
| In the recorder (NVR) | One point for several cameras, video stays on site | A limited number of channels with analytics, tied to the recorder maker, complex rules are rarely supported |
| On a server at the site | Works with cameras of different makes, the rule can be tuned to your process, frames do not leave the enterprise, no internet is needed to operate | Needs a server with a graphics card and space for it, and someone has to watch its condition |
| In the cloud | No hardware of your own, easy to bring several sites into one panel | Needs a stable link from every site, video leaves the enterprise, you need to know in which country and with which provider the frames are stored |
In production with dozens of cameras we put recognition on a server at the site and pass only the finished cards to the panel. For a chain of shops or several warehouses, your own cloud or our protected one is possible. How this choice changes the cost we covered in a separate guide on the price of video analytics.
Which tasks video analytics solves in different industries
Tasks differ a lot between industries, but almost all of them come down to three questions: someone broke a rule, something lies or stands in the wrong place, a process has stopped or dragged on.
| Industry | What a camera can control | Who receives the card |
|---|---|---|
| Food production, HACCP | Hair net, mask, beard cover, gloves in the zone, a box of product on the floor, hand washing at the entrance | Head of HACCP, process engineer |
| Occupational safety | Helmet, vest, a person in a danger zone, riding on a pallet truck, a pedestrian near a forklift | Safety engineer, shift supervisor |
| Warehouses and ramps | A truck on the ramp and its handling time, dock door idle time, cargo in an aisle, outsiders in the dispatch zone | Warehouse manager, logistician |
| Shops | Queue length at a till or counter, empty shelves, people in staff-only areas | Shift administrator, security |
| Agro | Outsiders on the grain floor and near machinery, a person near augers and in risk zones, vehicle movement at night | Agronomist, security |
| Construction | Helmet and vest, a person near a slab edge or under a crane, outsiders after working hours | Site manager, safety engineer |
A queue looks like a trifle until it is counted. Researchers from Columbia Business School photographed the deli section of a supermarket in Santiago every 30 minutes for 7 months (October 2008 to May 2009), counted the people in the queue on the images, and matched this with receipts. For the average customer, a queue growing from 5 to 10 people lowered the probability of a purchase as much as a 3.22% price increase would. Growth from 10 to 15 people acted like an 8.26% price increase. This is one study in one shop, but it shows well that a queue costs money even when nobody complains.
What from the table we already do in practice. At the meat processing plant where we ran the system, it checked the hair net, mask, beard cover, gloves, a box on the floor and riding on a pallet truck, and the head of HACCP went through the cards. Everything else in the table we set up for a specific site and always start in shadow mode first.
Why accuracy differs in a demo and on your site
In a demo the supplier shows clips where the camera hangs well, there is enough light and the person stands fully in the frame. On a production floor the camera looks down at an angle, people hide one another, there is steam from the washer, the sun shines in from a window, and a hair net almost merges with a white wall.
Ukrainian retail has already tested this. According to Valerii Bilyi, director of the security department at VARUS, in 2025 the chain tested solutions from six companies, and only one showed stable video analytics.
So accuracy has to be measured on your cameras. Here is how we do it and how we advise demanding it from anyone:
- The rule first runs in shadow mode: cards arrive, but nobody is penalised for them. The responsible person marks for a few weeks which cards are correct.
- Accuracy is counted by events. Twenty neighbouring frames of one person without a mask make one event. Otherwise a single case turns into twenty "hits".
- Next to any percentage there is the sample size, the period, and who decided. A percentage without these three things means nothing.
- Missed cases are looked at separately. For this, someone goes through a piece of ordinary recording and counts how many violations the system should have seen.
False cards have a cost too. Each one is opened, looked at and rejected by someone. If a hundred cards arrive per shift and ten are correct, within a week the person will stop opening them and the system becomes useless. So before going live we look not only at the share of correct cards but also at how many minutes a day they take from the responsible person. For more on the questions worth asking a supplier, see our guide on how to choose video analytics: 10 questions to ask.
Video analytics and personal data
Video in which a person can be identified is already personal data, and cameras at an enterprise need a legal basis, a purpose, a retention period and notice to employees. This applies to ordinary video surveillance too.
Most production tasks do not need to know who exactly is in the frame. To notice a person without gloves it is enough to see the hands. We do not recognise faces and do not build tasks on biometrics. What the law requires we covered in detail in a separate article on video analytics and personal data.
What to pay attention to in Ukraine in 2026
On 28 January 2026 the State Special Communications Service added another 13 programs to the list of prohibited software. Among them are Macroscop products (Macroscop, ML Macroscop, LS Macroscop, ST Macroscop) and products of the Axxon developer (Axxonsoft, Axxon Next, "Inter Lite", "Intelekt"). For state systems and critical information infrastructure the ban is mandatory. For an ordinary private business there was no direct obligation as of October 2026, but licences and support for such products are getting harder to obtain, so replacing the software part should be planned. More in a separate guide on replacing Macroscop and Axxon.
Cameras do not have to be replaced. The video analytics software and the camera are different things. Most IP cameras and recorders deliver video over RTSP, and many support ONVIF, so new analytics can be connected to the cameras already installed. New cameras are needed only where the existing ones do not see the required zone or the person in the frame is too small.
It is useful to know about ONVIF Profile M. It is the ONVIF profile for metadata and analytics events: object classification, counting, rules, and event delivery including via MQTT. If a camera supports it, its own analytics is easier to bring into one system with others. For tasks such as "is the person wearing a beard cover", the analytics built into cameras is usually not enough, and then processing runs on a server.
Think about power separately. During power outages cameras on backup power may keep working while the server or the communication link does not. Processing at the site helps here: it does not depend on the internet, and the server must be connected to the same uninterruptible power source as the cameras.
Frequently asked questions about video analytics
Do we need new cameras?
Most often no. The system takes video from your existing IP cameras or recorder. New cameras are needed where the zone is not visible or the person in the frame is too small.
How is video analytics different from a "smart camera"?
A "smart camera" has a built-in set of functions from the maker: line crossing, a person in a zone, sometimes plate numbers. Video analytics on a server works with cameras of different makes and can check rules built for your process.
Does the system work without the internet?
If processing runs on a server at the site, yes. The internet is needed only to view cards from outside the enterprise.
Does the system recognise employees' faces?
We do not do this. To control protective equipment, boxes or queues there is no need to know who exactly is in the frame.
How many false alerts will there be?
Nobody can honestly say in advance. The number depends on the camera, the light and the rule. That is exactly what shadow mode is for: within a few weeks you see the real number of cards and the share of correct ones on your events.
How long does launch take?
Connecting cameras is usually the fastest part. Shadow mode takes longer, and during it the rule is tuned to your site. Tasks the system does not handle yet need training on your frames, and we agree the target for them with you.
What should we do if Macroscop or Axxon is installed now?
They were added to the list of prohibited software on 28.01.2026. Replacing them is mandatory for state systems and critical infrastructure; a private business should plan a replacement. Cameras and recorders can usually stay, and the software part is replaced with another that takes video over RTSP or ONVIF.
Where to start
- Write down 2-3 situations that cost you the most: an auditor's fine, an injury, a write-off, a queue, a theft. Start with one.
- Check whether at least one existing camera sees this zone and delivers video over RTSP. Your IT specialist can check this.
- Decide who will go through the cards and how much time a day they can spend on it.
- Ask the supplier for a few weeks of shadow mode on your cameras and for accuracy by events with the sample size, before you agree to a live launch.
To compare vendors, see how to choose video analytics: 10 questions to ask. We can connect to your existing cameras and start with shadow mode on one rule. Contact details are on the Contacts page.
Where to apply this
- Food safety and hygiene
- Occupational safety
- Retail and supermarkets
- Dish pass time in a restaurant: what to count and how a camera does it
- Restaurant kitchen control with a camera: sanitation, HACCP and what the camera checks
- Video analytics for a restaurant: what it does in the kitchen, at the pass and in the dining room
Sources
- dev.ua. Another 13 programs added to the prohibited software list (State Special Communications update of 28.01.2026), 2026
- Retailers.ua. Ukrainian retail tests AI cameras: why business is in no hurry to adopt them (VARUS, 6 suppliers in 2025)
- Lu Y., Musalem A., Olivares M., Schilkrut A. Measuring the Effect of Queues on Customer Purchases, Columbia Business School, 2011
- ONVIF. Profile M: metadata and events for analytics