← Back to the blog
Product

Object detection with cameras: what changed in 2026 and whether it fits your business

Computer vision models are lighter and now run on simple devices. Technology is no longer the biggest obstacle. The hard decisions are now elsewhere, from the right images to licensing and data.

Camera and a detection box around a suitcase in dark lines on a light background, beside the glass form with a blue core

For years, getting a camera to "see" objects looked like a project for heavy industry, with expensive servers and months of research.

That is changing, above all in where the model can run.

What has changed

In January 2026, Ultralytics launched YOLO26, a new generation of one of the most widely used model families for object detection. The company presents it as designed to run on the device itself, and cites up to 43% faster CPU performance than the previous version, in one specific test.

It pays to look beyond the announcement. An independent study, using aquaculture images, found YOLO26n to be the fastest on a Raspberry Pi 5, a small, low-cost computer. The difference in accuracy between generations, however, was small in that case.

The practical takeaway is that running close to the camera is more viable, and the result still depends on your environment.

Flow in which the camera image passes through the model that detects the object, then the business rule, and becomes an alert, a count or a record in the system

Where detection works well

The best scenario has a clearly defined object and a clear action when it appears:

  • Suitcases and backpacks in airports and terminals, to follow items left standing.
  • Vehicle number plates at gates and yards, to log entries and exits.
  • Phones in classrooms, detecting the device and not who is using it.
  • Counting products and packages in a warehouse, dock or shelf.

What to assess before you start

  1. What should happen when the object appears? Without an alert, a count or a record in the system, the project remains a demo.
  2. Are the images from the real environment? Light, angle and distance change the result a great deal. The model should be trained on images from your operation.
  3. Which error costs more? Sometimes a miss is worst, sometimes a false alarm. Agree the target before the pilot.
  4. Does the licence allow your use? YOLO26 comes under the open AGPL-3.0 licence, which requires opening up the project, or an enterprise licence for closed commercial use. Account for this from the start.
  5. Where will the model run? On the device, on a local server or in the cloud. The closer to the camera, the fewer images travel.

Cameras call for care with data

A number plate can lead to the vehicle's owner, so treat it as personal data. Data about children, as in a school, needs stronger protection under data protection laws, including the GDPR in Europe.

Good practice includes setting out the purpose in writing, keeping the minimum for the shortest time, blurring faces in stored images and limiting who can access the system. Speak to your legal adviser before installing.

Start with a pilot

In computer vision and object detection, we start by defining what needs to be seen and the acceptance criteria. The pilot measures correct detections and false alarms in the field, and the result reaches your system as an alert, a count or a record.