Knowledge Base
Planning7 min read

Cameras as Sensors: When Vision Beats a Sensor, and When It Does Not

DBR77 IoT Team · Published

A camera beats a sensor where there is nothing to read from: manual assembly, packing, the welding bay or the press from 1994, where retrofitting sensors costs more than the station is worth. Where a machine has a PLC, the PLC signal wins, because it is more accurate and cheaper.

Below: where the real data gap sits, what a camera measures at a manual station, how a time study changes, and where vision is the wrong tool.

Where the real data gap is

CNC machines are rarely the problem. They have PLCs, and in many plants they already report to the MES.

The gap is manual assembly, packing, the welding bay and the old press. There is no controller to read, so these stations run on paper sheets, estimates and memory.

That gap matters for OEE. Vorne notes that most companies do not accurately track small stops and slow cycles. At a manual station they are not tracked at all. If your oldest machines do have some signals, our guide on connecting old machines without a PLC covers the sensor route first.

How a camera becomes a data source

DBR77 adds a Vision AI layer to IRIS, the DBR77 plant operating system, as one more data source. The path is the same as for controller telemetry: camera, edge GPU, MQTT, IRIS.

Because a vision event lands in the same system as machine data, it can trigger the same actions:

  • slow down an AMR fleet when a person enters a zone;
  • open a work order in the CMMS;
  • log a near-miss for HSE;
  • feed OEE with cycles and piece counts.

Processing stays on site. There is no video recording by design: the edge box turns images into events and sends only the events on. Our article on edge vs cloud in manufacturing explains why that split matters.

A pilot is small: 2 to 4 cameras, one edge box and a PoE switch. It is installed in a day, with no changes to the machines. DBR77 states that a camera for one station can cost around $50, and that calibration is the most expensive step.

See Vision AI on real stations

The Vision AI webinar covers operator-trained quality checks, tagless asset tracking and proximity alerts, all processed on a local edge device.

Register for the webinar

Time study with a camera instead of a stopwatch

Today an engineer stands at the station with a stopwatch and a notepad, records thirty cycles and takes an average. It costs a day. And it samples one day, one shift and one operator who knows they are being measured.

A systematic review in the Journal of Clinical Epidemiology found that being studied does affect the behaviour being studied. A stopwatch study can easily capture the best version of the job.

With a camera, the steps are different:

  1. Draw the work zone on the camera image with a mouse.
  2. Define a cycle as one stay of the operator in the zone. Debouncing stops a short step out from counting as two cycles.
  3. Read the output: cycle count, last cycle, average, minimum, maximum and presence percentage over the last 10 minutes.
  4. Add WIP tracking with printed ArUco markers, square fiducial markers with an ID, on parts or pallets. The camera counts what enters and leaves the zone.

You do not need to mount a camera to start. Upload a phone video of a station and measure it in minutes, then decide whether it deserves a permanent camera.

What a permanent camera tells you

Once the camera stays, the time study never ends. On a typical manual assembly cell, for example, an engineer sees:

  • cycle time against takt, every cycle, every shift;
  • variability per shift and per operator, which shows where standard work or training needs attention;
  • piece counts that feed OEE performance;
  • micro-stops with the cause visible: no material, no operator, waiting for a forklift.

The last item is the one stopwatch studies miss completely. A sensor tells you a station stopped. A camera also shows that the operator was standing there with an empty bin. To act on it, that event still needs context such as order, shift and product; see why machine data is useless without context.

When a sensor or PLC still wins

The line DBR77 holds is simple: where the machine has a PLC, the PLC signal wins. It is more accurate and cheaper than any camera.

| Station type | Best data source | Why | |---|---|---| | CNC or modern machine with PLC | PLC signal (OPC UA, Modbus) | Exact states and counts, nothing to calibrate | | Old machine with usable signals | Edge device with clamp-on sensors | Current or stack light gives reliable run and stop states | | Manual assembly or packing | Camera with Vision AI | No controller; the work is people moving parts | | Welding bay or old press with no signals | Camera with Vision AI | Sensor retrofit costs more than the station is worth | | Visual quality check | Camera with operator-trained model | Operators train it with OK/NOK buttons from good parts | | Micron-level tolerances | Metrology equipment | A camera does not replace a gauge | | Machine guarding | Certified safety system | Vision AI is a measurement layer only |

Operator-trained quality checks work well for visible defects, but micron tolerances still need metrology. And Vision AI is not a certified safety system. Safety functions on machinery are designed to standards such as ISO 13849-1; a zone alert that slows an AMR fleet adds information on top of them.

What about privacy and employee monitoring?

Cameras at workstations raise fair questions from employees. The DBR77 design answers part of it: processing on site, no video recording, and no face recognition. The system outputs cycles and events.

Workplace monitoring rules may still apply when a camera covers people. In Poland, art. 22² of the Labour Code allows video monitoring for purposes that include production control. The data protection authority UODO summarises the duties: employees must be informed no later than two weeks before launch, the purpose and scope must be defined, and recordings may be kept for no more than three months. Check the rules in your country and involve employee representatives before the first camera goes up.

FAQ

Can I monitor a machine without a PLC?

Yes. If the machine has usable signals, an edge device with clamp-on sensors reads them. If there is nothing to read, or the station is manual, a camera measures cycles, presence and piece counts instead.

How accurate is a camera-based time study?

It measures every cycle, across all shifts, so the sample is far larger than thirty stopwatch readings. Define the work zone carefully and check the first results against a short manual count.

Does the system record video of employees?

No. Images are processed on the edge box on site and turned into events. There is no video recording by design.

What does a Vision AI pilot need?

Two to four cameras, one edge box and a PoE switch. Installation takes a day and the machines are not changed. You can test a station first with a phone video.

Conclusion

Use the PLC where there is one. Use a camera where there is nothing to read, especially at manual stations where a stopwatch has been the only tool. Start with a phone video, measure one station and decide from the data whether it earns a permanent camera.

Add vision to your plant data

See how IRIS combines camera events with machine data, CMMS work orders and OEE in one system.

Explore IRIS

Sources

Knowledge BaseBook a Demo