Computer Vision Instead of Operators: How PPE Compliance is Checked on Ship Decks Via Cameras

The customer commissioned the development of a system for automatic monitoring of personal protective equipment (PPE) compliance among the crew on the open deck of a vessel, replacing manual visual control by camera monitoring with a violation detection system.

What Was the Problem with Manual Control

In work zones on a ship's deck, the crew must be in hard hats, life jackets and flotation suits. Formally, this safety requirement is monitored visually — an operator watches recordings or live broadcasts from cameras. In practice, such control does not scale: one person is physically unable to cover all video streams during a shift, especially when there are many cameras and work zones on the vessel, and the operator's attention inevitably becomes scattered.

• Object of control — open ship deck, crew work zones
• Mandatory PPE — hard hats, life jackets, gloves, flotation suits
• Problem of manual control — the operator cannot physically monitor all cameras during a shift
• Task — automatically detect PPE violations from the video stream
• Key phase — converting a stream of raw detections into a confirmed violation

Why Single Detections Are Not Enough

The task description directly indicates that a single detection of an object in a frame — for example, detection of a person without a hard hat — is not enough to register this as a real violation. A system working on individual frames will inevitably stumble over false positives: an accidental camera angle, partial occlusion of the figure, a moment when a person removes and immediately puts on a life jacket. This is why the task description separately highlights the transition from "stream of detections" to "confirmed violation" — that is, the system needs logic that separates isolated artifacts of recognition from a sustained, recurring fact of missing PPE on a specific person in a work zone.

Such an approach is typical for industrial computer vision on real-world objects: a single detection is a signal, not a ready verdict. For the verdict to be suitable for real-world use — for example, for alerting the responsible safety officer on the vessel — the system must be built so that it aggregates detections over time and by specific person or zone before considering a case a violation.

What This Means

The task of monitoring PPE on a vessel is a characteristic example of how computer vision replaces manual observation not where a single detection is needed, but where a system is needed that can distinguish noise from real events. For industries like maritime transport, where manual safety control physically cannot scale to dozens of cameras and work zones, this kind of automation becomes not just a convenience, but a way to cover blind spots that an operator objectively cannot monitor.