In short
A camera captures pixels. Computer vision turns those pixels into useful information: there is a person here, this box is damaged, this label reads “Batch 42”. Most modern computer vision uses neural networks trained on large numbers of labelled images. Newer vision-language models can also describe a picture or answer questions about it in plain English.
The main tasks
| Classification | What is in this image? For example, a pass or fail on a finished part. |
|---|---|
| Object detection | What is in the image, and where? Boxes drawn around each person, vehicle or product. |
| Segmentation | Exactly which pixels belong to each object, useful for measuring areas or finding the edge of a defect. |
| Text recognition (OCR) | Reading printed or handwritten text from labels, documents and signs. |
| Tracking | Following objects or people across video frames to count them or measure movement. |
| Pose estimation | Working out the position of a body or object, used in ergonomics and robotics. |
How businesses use it
Quality inspection is one of the longest-established uses. Cameras on a production line check parts, labels, fill levels and packaging faster and more consistently than a person can over a long shift.
In warehouses and shops, vision counts stock on shelves, reads barcodes and labels, and checks that orders are packed correctly. In construction and facilities, it can monitor progress from site photos, read meters and flag hazards. Automatic number plate recognition (ANPR) runs car park barriers and site access across the UK.
Document processing relies on it too. Reading invoices, forms and delivery notes starts with recognising text and layout in an image.
Robots depend on it. A robot arm picking parts from a bin, an autonomous floor cleaner and a humanoid at an event all need to see. Jimmy, our Unitree R1 Air humanoid, runs object and person detection on his camera feed at roughly 120 milliseconds per frame, and recognises the faces of people he has been introduced to.
What makes a project work
Consistent conditions. Lighting, camera position and background have a large effect on accuracy. A fixed camera over a conveyor with controlled lighting is far easier than a phone photo taken on a muddy site.
Representative examples. A system trained or tested only on perfect images will struggle with real ones. Collect examples of the actual problem, including the rare defects you most need to catch.
Clear limits. Know what the system cannot do. Jimmy has no depth sensor, for example, so his software makes no claim to detect obstacles from vision. A system that knows its limits is safer than one that guesses.
Cameras, people and the law
If cameras capture identifiable people, UK GDPR applies. You need a lawful basis, clear signage and notices, a sensible retention period and secure storage, and the ICO publishes guidance on video surveillance. Facial recognition used to identify individuals involves biometric data, which is special category data under UK GDPR and needs extra justification and safeguards.
Where possible, design for privacy: point cameras at products instead of people, blur faces automatically, or process images on the device and keep only the result, such as a count.
Tell staff before cameras go in, explain what they are for, and decide in advance who can view footage and for what purpose.
When this matters
Relevant if staff spend time checking, counting or reading things by eye. New Era AI can help you assess whether a vision project is practical as part of AI strategy work. Jimmy shows live object and person detection running on our own hardware.
Questions
Do we need special cameras?
Not always. Many projects start with ordinary IP cameras or phone photos. Fast production lines, small defects or poor lighting may need industrial cameras and controlled lighting.
How accurate is computer vision?
It depends on the task and conditions. Test it on your own images in your own conditions before relying on any accuracy figure from a supplier.
Is facial recognition legal for a business in the UK?
It is not banned, but using it to identify people involves biometric data, which is special category data under UK GDPR. It needs a strong justification, a DPIA and careful safeguards. Take advice before using it.