AI detection in a factory: what the camera sees and where to start
Author: DuoVision team
For: plant directors, heads of HACCP and safety, production and logistics managers who are considering computer vision in production
Summary
- AI detection in a factory is a program that watches video from the cameras you already have and notices violations by rules you set: a missing hair cover or gloves, a crate on the floor, riding a pallet jack. For docks and warehouses we set rules up per site.
- DuoVision, a Ukrainian video analytics company, runs this on existing cameras, with no new cameras and no new video surveillance system. Frames are processed on a server at your site.
- Every trigger becomes a card with a frame, and a responsible person, for example the head of HACCP, decides whether it is a violation.
- The best start is a 4 to 6 week pilot on one zone in shadow mode, connecting only the rules you actually need.
For: plant directors, heads of HACCP and safety, production and logistics managers who are considering computer vision in production
AI detection in a factory in plain words
It is a way for a program to watch camera video and understand what happens in the frame: where a person is, what they wear, where a crate stands, whether a trolley is moving. This field is called computer vision. The program watches camera video against the rules you set, and the decision stays with a person.
You set a rule, for example "everyone in the processing room wears a hair cover and mask". When the rule is broken, the program creates a card with a frame. The head of HACCP or the safety team sees the card, judges it, and records the decision in a log. Video surveillance and computer vision together then produce a list of specific cases instead of just an archive of recordings. For the general picture see "Video analytics: what it is, how it works and what it gives a business".
The program does not replace a person. It picks out of thousands of hours of video the few minutes worth looking at.
What AI detection sees in a factory
Hygiene and PPE
At the meat processing plant where we ran the system, rules checked the hair cover or net, mask, beard cover, and gloves. More: PPE control by cameras and glove control. For other production lines, such as dairies or bakeries, we set rules up per site and start in shadow mode.
Crates on the floor
The rule sees a crate standing directly on the floor and ignores one on a pallet, trolley, or stand. There is a separate article: crates on the floor.
Pallet jacks and forklifts
We checked a person riding a pallet jack with a platform at the plant: riding a pallet jack. Forklifts moving near pedestrians in a warehouse or at a dock are set up per site: forklifts and pedestrians.
Docks
For docks (for example, truck dwell time) we set rules up per site and start in shadow mode. More in the article about the loading dock.
Vehicle plates
We recognize vehicle plates at the gate: the system records plate, time, and a photo. It needs a camera aimed at the traffic lane. Details: plates at warehouse and elevator gates and tracking vehicles on plant territory.
What suits you depends on your production. So every rule beyond those tested at the plant first runs in shadow mode and is switched on for real only afterwards.
What AI detection does not do
- It does not recognize faces. It checks whether a hair cover is on a head, not whose head it is.
- It does not hear sound or weigh cargo. A camera provides only the picture.
- It does not see what is hidden: behind shelving, under a trolley, outside the frame. The rule then marks the card "undetermined" instead of guessing.
- It does not promise absolute accuracy. Accuracy depends on the camera, light, and how the rule is defined. It is checked at your site.
- It does not decide for a person. A card is a hint, and the specialist decides.
On personal data: video analytics and personal data.
How it is set up on site
We work on cameras that are already installed. Frames are processed on a server at your site. From you we need access to the camera streams and a responsible person who reviews the cards. ONVIF Profile S describes how a camera passes video to a client, for example video management software. More on camera requirements in video surveillance cameras with artificial intelligence.
The rule creates a card with a frame, a person reviews it, and the decision goes into the log.
Where to start: a pilot on one zone
A launch across the whole plant at once is almost never needed. We suggest a 4 to 6 week pilot on one zone, for example a processing room entrance, a deboning line, a dock, or a gate. In practice:
- Survey. We look at the cameras in the zone, check what they show, and write down with you what counts as a violation and what counts as normal.
- Shadow mode. The rules run, cards go to the responsible person, and nobody is punished for them. The person marks cards right or false.
- Tuning. Meanwhile we study the false cards and fix the rule, the zone, or the angle.
- Result. You see how many violations were found, how many cards were extra, and whether the rule is worth switching on and extending to other zones.
We described shadow mode in the article on the meat plant launch, which shows how the rules changed after the first weeks.
The benefit is plain. You buy no new cameras and no new video surveillance system, because we work on those already installed. You connect only the rules you need. A pilot on one zone shows the result before you decide to expand.
What it costs
An exact figure can only be given after a site survey. Cost depends on the number of cameras in the zone and the number of rules. There is usually no extra spending on new cameras or a new video surveillance system, because we use the existing ones.
The overall picture is in what video analytics costs and when it pays off. You can write to us with a description of your zone, and we will tell you what it needs.
Frequently asked questions
Do I need new cameras?
Usually not. We work on cameras that are already installed. A new camera is needed where the existing one does not see the zone or a person in the frame is too small.
Where are frames processed?
On a server at your site.
Who decides it is a violation?
A responsible person on your side. The system shows a card with a frame.
Do you recognize faces?
No. We recognize vehicle plates.
Has this been proven in practice?
Yes, at a meat processing plant where we ran the system. For other production lines, a shadow mode pilot shows the result.
What is computer vision in production?
Programs that understand camera images: they see people, clothing, crates, and vehicles. In production they serve hygiene control, workplace safety, and logistics.
How long does a pilot last?
Usually 4 to 6 weeks for one zone.
First step
- Choose the zone where violations cost you most.
- Describe what counts as a violation there.
- Write to us through the contact page and attach a few camera frames.
- Run a pilot in shadow mode and judge the result on your own frames.
On choosing a vendor, read "How to choose video analytics: 10 questions to ask".
Where to apply this
- Meat processing
- Food manufacturing
- Food safety and hygiene
- Production processes
- 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