How Accurate Is a Camera Part Count for OEE? What Affects It and How to Test It

Key takeaways

• View and job decide accuracy. A camera count feeds OEE (overall equipment effectiveness) best on one clear question: how many units passed this point?
• Errors come from the field of view, light changes, people or guards in the way, look-alike product and a camera that has moved.
• A count can drift for weeks unseen unless you check it against a hand count.
• Where a machine signal exists, it times the stops. The camera counts the parts.

By the numbers

• 28% of large US manufacturers rank vision systems as a first or second investment priority for the next 24 months. (Deloitte, 2025)
• Automated optical inspection is the leading industrial AI use case, with a share of about 11% across the 48 use cases IoT Analytics tracks. (IoT Analytics, 2025)
• Analytical AI and machine learning are embedded in nearly 62% of Lighthouse sites' top five use cases, and generative AI in 23%, up from 9% in 2024. (World Economic Forum, 2026)

A camera part count is accurate when the camera has a clear, steady view of units passing one point. Accuracy drops when people, guards, changing light or look-alike product get in the way. Published accuracy figures are thin and mostly self-reported by vendors, so the number that counts is your own test.

Fabrico is a cloud-based OEE and light MES (manufacturing execution system) platform. Stops come from the machine signal, and the operator adds the reason. The AI camera (an add-on) counts parts. This guide quotes no accuracy figure for any camera, ours included: it shows what to test.

What makes a camera count miss units or add extra ones?

A camera sees one field of view, in one range of light, for units someone defined. Watch five things.

• Field of view. A unit that leaves the frame early, or a lane the camera cannot see, is not counted.
• Light changes. Daylight from a dock door, other lamps on nights or glare on shrink film can fool it.
• Occlusion. Hands, an operator leaning in, an open guard door or a parked pallet can hide the product.
• Look-alike product. Clear bottles on a shiny belt, or packs that touch and read as one, make counts harder.
• Camera drift. A camera knocked during cleaning, or a guide rail moved at a changeover, shifts the view.

The job matters too. "Is product moving?" is easier than "how many packs passed?", and both are far easier than "why did it stop?". A camera does not answer that last question: a person opens the video for that minute and looks, and the operator adds the reason.

See our guide to placing and lighting a line camera. For the wider setup, read about camera based production monitoring.

What does a wrong part count cost?

• Count too high: a pack is counted twice, or a reject that was pulled off still counts. OEE looks better than the line is.
• Count too low: packs that touch read as one, or a lane is hidden. Output looks worse than it is, and nobody knows why.

So test the count against a hand count at normal speed and at the busiest moments of the shift. A wrong count can hide lost output.

That loss can be large. Siemens' True Cost of Downtime 2024 report puts a lost hour in FMCG (fast-moving consumer goods) at about $36,000 for large plants.

And operators who correct a count by hand all shift soon stop trusting it.

Is a camera count as accurate as a PLC counter?

Not always, and it does not need to be. Where a machine has a reliable counter, it is usually exact for its own cycles. A camera counts what passes through its view.

So the machine signal is the clock: where a reliable signal exists, its times go on the record. The camera count covers stations with no usable counter, and checks the counter where one exists.

They can disagree for a good reason: a starved machine can show a run signal while nothing moves, and only the count shows the output is not there.

Keep the PLC, cameras and servers on one network time. Our comparison of computer vision and PLC data for OEE shows what each source does well.

How do you test a camera count on one station?

Run the camera count next to a hand count on one station.

1. Pick one station where the loss is and the view is clear.
2. Write down what you count, such as "one case passes the outfeed point", agreed by someone who runs the line.
3. Keep the current counting as it is. Do not change how people count during the test.
4. Hand count fixed windows, such as 10 minutes of output, and compare each with the camera count for the same window.
5. Line up the numbers for each window: the hand count, the camera count and the machine counter where one exists.
6. Cover real conditions: day and night shifts, a changeover, a clean and each product family.

Agree in advance which gap means "add the next station". The test is long enough when it has covered those conditions and a few dozen counting windows. For the wider plan, read how to run a computer vision OEE pilot, where the camera is an add-on.

What does a worked example of the test show?

Sample line (illustrative numbers, not a customer). A camera counts cases at the outfeed of a case packer on a bottling line. The case packer has a PLC counter. The test runs for 10 shifts, with 2 hand counted windows of 10 minutes a shift.

Method note: Sample line, illustrative numbers, not a customer. Every result above comes from the inputs listed and the formula shown with it, so you can repeat the calculation with your own line's numbers.

• Camera count 11,940 against hand count 12,000: 60 fewer, or 60 / 12,000 = 0.5% low.
• 17 + 3 = 20 windows. The 3 night windows lost 45 cases, 15 each, or 15 / 600 = 2.5% of a window.
• The other 17 windows lost 60 minus 45 = 15 cases in all, under 1 a window.
• The 3 bad windows had one cause: the dock door was open on nights and the light changed.
• A hand count of those windows finds it. A count checked once on a day shift would not.

Stop times in this test came from the machine signal, not the camera. The test checks counts only.

Result: the day counts pass the rule the team set, within 1%. The team shades the night light, repeats the 3 night windows and keeps the PLC counter as the record where one exists.

What should you do with the test result?

Each line of the result points to a different fix.

• A gap you cannot explain: see how computer vision feeds each OEE loss (the camera is an add-on), then check the view and the light.
• Misses with one cause: fix the view or the light, then check again.
• Counts too high: tighten what counts as a unit, or the field of view.
• The hand count and the PLC counter disagree: a counting problem at the machine, not the camera.

Add the next station when this one passes the rule you agreed. Fabrico starts with a pilot on one line. With the AI camera (an add-on), you open the camera video for that minute: see the computer vision add-on for OEE.

Ask any vendor, us included, how an accuracy figure was measured: which station, how many units, who counted them by hand. There is no free plan and there are no tiers: each plant gets a quote.

Proof from a named plant: Ficosota (also a Fabrico investor) raised OEE from 53.1% to 59.6% on 16 lines, June to September 2026, with monthly good output up 23.2%. Read the Ficosota case.

Toyota calls jidoka "automation with a human touch": the machine stops on a problem, and a person decides what happens next. (Toyota, Toyota Production System)

Frequently asked questions

Is computer vision accurate enough for OEE?

It can be, for clear jobs: how many units pass one point. It is weaker when the view is blocked, the light changes or products look alike. Published figures are thin, so run your own test on one station against a hand count.

Why do vendors quote such different accuracy numbers?

There is no shared test. Each figure comes from a different station, product and lighting, and from a different definition of a count. Many are self-reported. Ask how many units were checked, who counted them by hand, and whether nights and changeovers were included.

Can a camera replace the PLC for OEE?

Not where a reliable machine signal exists. The signal is the better clock for stop times, and a PLC counter is exact for its own cycles. The camera adds a count where there is no usable counter, and a check on one where there is. Many plants run both.

Does the camera decide why the line stopped?

No. The machine signal logs the stop and the operator adds the reason. If it is unclear, a person can open the camera video for that minute and look. The camera counts parts; it does not track people: no faces, no names, no scores per operator.