Машинное восприятие в промышленных и логистических средах

Computer vision and sensor-stream processing that has to work on a shop floor or in a warehouse rather than in a laboratory: changing light, noisy data, limited compute, hard reliability requirements. IPIIR's line examines which models and pipelines survive those conditions on real production lines.

Because error costs are asymmetric in production: a false alarm halts the flow, while a missed defect reaches the customer. The line selects measures beyond the benchmark — by error class, capture conditions and consequences for the process — and assesses how well a model adapts to one particular plant.

As one link in an existing pipeline: the model's signal enters a SCADA stack or an MES system and becomes an action — halting the line, rejecting a part, writing a log entry. The line studies such integrations, including where a human belongs in the decision loop and how the system behaves when a camera or a channel fails.

In logistics: recognising cargo, tracking movements and reading labels in field conditions. In occupational safety: monitoring personal protective equipment, detecting violations and preventing incidents. Both groups of tasks need sector-specific datasets, so the line also works on collecting and labelling them, including work with small samples.