- Home
- Resources
- Video surveillance cameras with artificial intelligence: which ones work and whether you need new ones
Video surveillance cameras with artificial intelligence: which ones work and whether you need new ones
Author: DuoVision team
For: owners and managers of plants, warehouses, fuel stations, and retail sites, and IT directors who wonder whether they must replace cameras for AI video surveillance
Summary
- AI video surveillance usually works on the cameras you already have: an IP camera that streams video over RTSP or ONVIF, with a clear picture of the zone where a rule is needed.
- DuoVision, a Ukrainian video analytics company, connects to such cameras and does not require a new video surveillance system. Frames are processed on a server at your site.
- A new camera is needed where the old one does not see the zone, people in the frame are too small, the picture is empty at night, or the camera is analog with no network output.
- Start with a pilot on one zone: we check a few of your cameras and run one rule in shadow mode to see the result on real frames.
For: owners and managers of plants, warehouses, fuel stations, and retail sites, and IT directors who wonder whether they must replace cameras for AI video surveillance
Which camera you need for video surveillance with artificial intelligence
AI video surveillance does not need a special "smart" camera. It needs a camera that sends a clear picture over the network. The camera only films and streams, and recognition happens separately from it. A program that reads the stream does the recognition. Names such as "AI cameras" or "smart cameras" usually mean cameras with built-in analytics, and for most tasks they are optional.
Most modern IP cameras and recorders stream video over RTSP. Axis documentation, for example, says its cameras serve video over RTSP and that stream parameters such as resolution and compression are set in the stream address. Many cameras also support ONVIF. Profile S describes how a camera passes video to a client, for example video management software.
If you already have video surveillance, you most likely have suitable cameras. For the basics of video analytics see "Video analytics: what it is, how it works and what it gives a business".
You mostly do not need to buy new cameras
This is the question newcomers ask first. At the plant where we ran the system, we installed no new cameras. We worked with the ones already hanging in the processing rooms and at the docks. More in "Video monitoring of hygiene in a meat plant".
For you it means:
- no set of new cameras just for analytics;
- no need to replace your video surveillance system: the analytics run alongside it and take video from the same cameras;
- you connect only the rules you need;
- frames are processed on a server at your site.
If you use Macroscop or Axxon and have to replace the software, the cameras usually stay. Details are in the Ukrainian article "Replacing Macroscop and Axxon".
What to look at in your cameras
Network output and protocol
The camera must be an IP camera that streams over RTSP or ONVIF. Analog cameras without a network output connect through a recorder or a signal converter, and picture quality then depends on them. If a camera gives video only through the maker's own program, find that out before work starts.
Angle
This matters most and is underrated most often. A rule sees only what is in the frame. Checking a hair cover needs the head in view, a crate on the floor needs the floor around the crate, a pallet jack needs the aisle. A ceiling camera looking almost straight down counts people well but shows faces and hands poorly. A camera high in a corner sees a lot, but people look small.
Resolution
The smaller the detail, the more pixels you need. Seeing that there is a person is easier than seeing whether they wear a mask or a beard cover. For license plates the camera needs enough pixels across the plate, so a separate camera at the gate is often used. See plates at warehouse and elevator gates. For other rules we do not rely on the number on a camera's spec sheet. We check whether the rule is visible on a real frame from your site.
The stream the camera gives
Many cameras have two streams: a main one in high quality and a lighter secondary one. Analytics normally use a stream of sufficient quality, and it must run steadily. Also check how many simultaneous connections the camera supports if your recording system already reads it.
Light
Backlight (a window or gate behind people), shiny surfaces, and sharp shadows make the task harder. A few frames at different times of day show it. Processing rooms with even light usually give fewer problems than a yard or a dock.
Night and infrared
Many cameras switch on infrared (IR) illumination at night and give a black and white picture. When it is enough to see a person, a vehicle, or movement, that often works. For fine detail such as clothing color or crate contents, an IR picture is worse. We therefore test rules that depend on color or small details separately by day and by night. IR glare on glass, water, and wet surfaces also gets in the way.
Checklist: check your cameras
You can do this in an hour with your IT person.
- List the cameras covering the zone where you want a rule: model, network address, where it hangs.
- Check that each is an IP camera and streams over RTSP or ONVIF. Opening the stream in an ordinary video player is enough.
- Open the live picture and check that what the rule must verify is visible: head, hands, floor, aisle, plate.
- Find frames from different times of day, including night and the worst light.
- Check that the picture is stable: no blur, no fogged lens, nothing such as shelving, doors, or hanging carcasses blocking the view.
- Check whether the camera covers the spots where people most often break the rule.
- Note whether cameras and the server have backup power for outages.
- Send us a few frames and the camera models. We will tell you whether they suit your rule.
When a new camera is needed after all
- Existing cameras do not see the zone, for example a door into the processing room or a particular table.
- A person or part in the frame is too small, and only a new camera can get closer.
- The camera is analog, has a poor signal, or does not record at the quality you need.
- At night the picture is empty, and you need the rule at night.
- For plate recognition at a gate, which needs a camera aimed at the traffic lane.
Install a new camera only in the zone where the pilot showed a problem, not in advance across the whole site.
What comes after the camera
A camera supplies video. To turn it into control you need a rule that describes the violation. When the rule triggers, the system creates a card with the frame. A responsible person, such as the head of HACCP or the safety officer, decides whether it is a real violation. The decision goes into a log, which lets you count violations and see trends.
A rule first runs in shadow mode. Cards arrive, but nobody is punished for them. After a few weeks you see how many cards are false and whether the rule is visible from your camera at all. Only then is it switched on for real. More on the approach: PPE control by cameras.
Frequently asked questions
Will my old camera work?
If it is an IP camera with RTSP or ONVIF and its picture shows the zone you need, yes. This is checked on a few frames and needs no purchase.
Do I need AI cameras or "smart cameras" with built-in analytics?
No. Recognition happens separately from the camera and reads its stream. Built-in camera analytics help with simple events; whether they are enough for a check such as "is a beard cover worn" depends on the camera and the task. ONVIF Profile M also exists for analytics metadata. If your cameras support it, that is a bonus, not a requirement.
Can I use cameras from several makers at once?
Yes, if each streams over RTSP or ONVIF.
How many cameras does a pilot need?
Usually a few cameras on one zone. The exact number depends on the zone and the rule. Cost depends on the number of cameras and rules and is best discussed after a site survey. For the general picture, see what video analytics costs.
Do you recognize faces?
No. We recognize vehicle plates.
Where are frames processed?
Video is processed on a server at your site.
Where to start
- Pick one zone and the one rule that matters most.
- Run the checklist above for the cameras in that zone.
- Send us frames and camera models through the contact page.
- Run a pilot in shadow mode and judge the result on your own frames.
On choosing a vendor, see "How to choose video analytics: 10 questions to ask".
Where to apply this
- Production processes
- Site security and vehicles
- Food safety and hygiene
- 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