Industrial computer vision

Computer vision converts an image into structured information: what is present, where it is, what state it is in. In an industrial setting that means reading assets, people, machinery, vehicles and zones from the cameras, drones, robots and embedded devices already deployed across a site.

The industrial qualifier is not marketing. A model evaluated on clean, well-lit, well-framed imagery is being evaluated on conditions that a working site does not provide. Sites have fog, glare, floodlight, night, dust, rain, vibration, occlusion and cameras mounted for recording rather than for analysis, and a model whose performance depends on the absence of those things has not been tested against its actual deployment environment.

Many industrial questions are not single detections but chains of them, and treating them as one step is where systems quietly fail. Verifying personal protective equipment is the clearest example: the question is not whether a helmet is present in the frame but whether this person is wearing one, which resolves as a chain, person, then head, then helmet, each step scoped by the one before it.

Structuring it that way is what makes the result auditable. Each link can be checked, and a failure is locatable rather than being an opaque disagreement with a single confident answer.

Fixed infrastructure is the part of the scene that should not change, which makes it the part where change is informative. A vessel, a valve skid, a panel, a pipe rack, a transmission tower: each has a normal visual state, and deviation from it, an open enclosure, a displaced component, an obstruction, a plume, is a finding precisely because the object is otherwise static.

Zones work the opposite way. A zone is defined by what is supposed to happen inside it, and the question is whether what is happening matches. Both are comparisons against an environment's own baseline, which is why models are specialised to a site rather than configured with thresholds.

Some industrial questions require an extent rather than a yes. Whether a fluid is present is a weaker question than how far it has propagated; whether spray is occurring is weaker than how much and where. Arkania has delivered detection with propagation measurement in oil and gas operations, and spray detection with measurement and control, a case several other vendors attempted before it was resolved.

Drone-based work extends the same principle to areas fixed cameras do not reach, including construction progress control against a checklist, applied on pipeline works.

The cameras on an industrial site were installed over many years, by different contractors, for recording. They are heterogeneous in make, age, resolution, mounting height and lighting, and they will not be replaced to suit an AI project.

Treating that as the starting condition rather than as a problem to be fixed is a design requirement. Integration with systems already in operation follows the same rule, SCADA, OPC UA, MQTT and M2M, so that what vision produces arrives in the control environment operators already use.

A computer vision model is only as good as its fit to the environment it watches, and environments change: a line is reconfigured, a camera is re-aimed, a yard is re-laid, a season changes the light. Specialisation to a site is therefore not a one-off setup step but a lifecycle property.

Arkania holds that lifecycle in-house, models designed, developed, trained and operated by Arkania, owned at code, architecture and lifecycle level. Re-specialising a model when the environment moves is part of operating the system rather than a new procurement.