Development of an AI system for PPE control and occupational safety

PPE Monitoring System: Where Manual Walkthroughs Break Down and What the Site Loses

Manual site walkthroughs are set up like this: once per shift, a foreman walks the area and records violations in a paper occupational safety log.

Between walkthroughs, the zone is not monitored at all — a hard hat is removed during trench work, a fall protection harness is clipped to an anchor as an afterthought, and no one will see it until the next check.

A 10–15 minute sample per shift does not reflect the real picture: a violation will make it into the report only if the foreman physically reached that point at that moment.

The log does not work as a carrier of evidence. Rows are filled in from memory at the end of the shift, signatures are collected retroactively, and there is no photo evidence.

When an incident is reviewed, the commission receives a contradictory set: the log says “no violations,” while the explanatory note says “worked without safety glasses for the third day.” PPE Provision Rules record the fact that protection was issued and used, but they do not confirm that the item was on the person at the exact moment of the operation — and in a disputed situation, linkage to time and place matters.

Then the cost comes into play. An inspector requests confirmation for a specific worker and area, and the site can show only handwritten rows without timestamps.

A dispute with the regulator is lost not because the process did not exist, but because the process cannot be confirmed. A separate point of failure is concealment: a foreman whose plan and bonus depend on the absence of violations has no incentive to write anything extra.

Each of these gaps is closed in its own way: PPE wear monitoring by zones instead of walkthroughs, machine timestamps instead of signatures, event export instead of rewriting the log. Let us break down how this works and what changes on the site side.

Implementation Effect: Shift Coverage, Response Time, and Failure Class

Sample checks produce statistics, not a picture: a once-per-shift walkthrough physically covers a few percent of observations, and a violation in a crew most often goes unnoticed.

Automated occupational safety monitoring changes the very principle — the stream from the site is analyzed in full, so every shift makes it into the report. Shift coverage becomes continuous: you see not an average for the area, but the specific shop, time, and workplace.

Response time stops depending on the walkthrough schedule. Previously, a violation went into the log and surfaced at the planning meeting or at the end of the shift; now the event reaches the responsible person within a minute — while the person is still at the dangerous point.

The class of response changes: not a review at the end of the week, but stopping a specific action on the spot.

Hence the eliminated class of violations — chronic ones that persisted for months: work without a fall protection harness, a rigger without a hard hat for an entire shift, unauthorized people in the area of vehicle movement.

The logic of checks is tied to Uniform Standard Norms for Issuance of PPE, so what is recorded is not “a person without a hard hat” but a deviation from the norm for a specific operation.

The reduction in injuries is not visible immediately: the shift is assessed using a rolling window, by the share of shifts without violations, and by the number of repeat triggers in one area.

The system does not cancel walkthroughs and work permits. It adds a continuous layer of monitoring, on top of which sample checks now go to specific addresses — to places where repeat violations are recorded, rather than at random across the shop.

Architecture Options: Cloud, Local Circuit, Hybrid

The deployment scheme is determined not by convenience but by security service requirements and the state of the communication channel at the site.

The camera video stream is the most sensitive circuit: you need to answer in advance where it may be stored, what will happen during an external channel outage, and who maintains the server side. Below are four schemes that are realistically chosen between during implementation.

In practice, they start with the scheme that the security service approves, and within it they select equipment for the number of cameras and archive depth.

Moving from a local circuit to a hybrid one later is cheaper than changing the entire architecture — that is why we design on-site deployment with headroom for such a transition.

Implementation Stages: From Pilot Zone to the Entire Shift

Implementation proceeds in waves: first a pilot zone, then the remaining areas. This way, errors in alert logic and settings are identified on a few cameras rather than across the entire site, and the production process does not stop.

1. Site survey. We walk the shift routes, mark places where entry into a danger zone is not controlled, and examine lighting, obstructions, and vehicle movement. The output is a list of areas and constraints that affect camera placement.
2. Monitoring zone scheme. We agree on where a hard hat, vest, glasses, or hearing protection is mandatory and which rule applies in each area. The document records the mapping of zones to violation types — alert logic and responsibility are built from it.
3. Installation and setup. We install cameras and the computing node, route power and network, and enable event-based clip recording. Here we also connect notifications for the duty shift and security.
4. Camera and threshold calibration. For each zone, we select the angle, detection area, and trigger threshold to remove false alarms from glare, reflections, and similar workwear. We test settings on archive recordings, comparing triggers with actual violations.
5. Trial operation. The system runs in notification mode without automatic sanctions: shifts get used to monitoring, while we collect statistics on false and missed triggers and retrain the model on real footage from the site.
6. System acceptance and handover. We run control scenarios, record agreed metrics, transfer proven parameters to the remaining zones, train operators, and hand over regulations, documentation, and access.

The transition to the next stage happens only after the previous result is approved: the zone scheme is signed off, calibration is accepted, and trial operation has delivered predictable accuracy. This is how the pilot zone becomes the reference for deploying across the entire site.

What We Hand Over: Zone Schemes, Trigger Rules, Procedures

We assemble the documentation package in parallel with implementation, not “after handover” — otherwise the PPE monitoring system remains a black box that only its author can configure.

As a result, the customer receives a set that new engineers and the occupational safety service can use independently: change thresholds, add zones, and investigate incidents without contacting us.

The composition is as follows: the technical part describes where and how the equipment is installed, while the organizational part describes who responds to an event and within what time frame. Below is exactly what we hand over.

• Camera placement scheme and zone marking — a site plan with viewpoint locations, blind spot coverage, and work zone boundaries, including entrances, stairs, warehouses, and high-risk areas.
• Trigger rules and thresholds — which events are considered violations, how many frames confirm the fact, what duration of presence in a zone without a hard hat or vest produces an event, and which interval is discarded as false.
• Role and access matrix — who sees the video stream, who sees only events, who closes an incident, who exports the archive; separation by area and shift.
• Response procedure — response times for events by class, escalation order, recording of the investigation result, and actions for repeat violations by one employee.
• Shift instructions — a short operator memo: how to distinguish a violation from interference, what to do during a channel failure, and where to report a false trigger.
• Integration passport — event export formats, fields and retention period, and an exchange scheme with existing access control and notification systems.
• Rule change procedure — how to add a zone or change a threshold independently, with a change log so settings do not diverge between sites.

Every artifact is tested on real shifts before handover: rules that produce noise are rewritten immediately. Describe your task — we will come back with an estimate and a documentation scope plan for your site.

Case Study: How a Manufacturing Facility Eliminated Manual Walkthroughs

Metal structures manufacturing, three shifts, five shops. PPE monitoring relied on walkthroughs: a foreman walked the route once per shift, recorded violations in a paper log, and the review was postponed to the next planning meeting.

By that time, the person had already removed the hard hat and moved to another area — the violation record had no consequences. This implementation case study is Northern Metal Structures Plant.

Cameras were installed at shop entrances and in high-risk zones, and detection was configured for specific signs: hard hat, safety glasses, high-visibility vest, and fall protection harness during work at height.

The confidence threshold was raised so that welding glare and metal reflections do not go into notifications: the responsible person receives a confirmed event with a frame and timestamp, not a stream of alarms. Manual walkthroughs were abandoned completely — the foreman is busy with work, not with the route.

Metrics Before and After

When a zone is under observation continuously rather than at the moment of a walkthrough, behavior also changes: there are fewer violations without additional reviews.

“Before, we discussed what could no longer be changed; now we discuss what is visible in the frame,” says the head of the occupational safety service.

What to Do If Cameras Are Already Installed but There Is No PPE Analytics?

The existing surveillance fleet is almost always reused — replacing cameras makes sense only where the angle physically does not allow PPE to be seen. Work begins with an inventory: a map of points, viewing direction, obstructions, and zones that the camera already sees reliably.

The output is a suitability map where each point is marked as ready, requiring improvement, or useless for monitoring.

Angle suitability matters more than model quality. A hard hat is reliably read by a camera mounted above head level; a vest is read by its contrasting stripes if a person is walking toward or across the frame.

A point looking along the flow shows the backs of heads: movement is visible, but the presence of protection cannot be confirmed.

That is why we look not at “a camera exists” but at geometry — mounting height, tilt angle, distance to the work zone, and whether a person fits entirely in the frame.

Light changes the task more than it seems: backlight in an opening, night illumination, dust, and fog disrupt recognition of vest color.

The archive must be representative — at least two to three weeks, including night shifts and bad weather: we use it to check detection behavior and select trigger thresholds.

Improvement is almost always easier than replacement: change the lens, rehang on a bracket at the right angle, add a local spotlight, cover blind zones with one or two cameras.

Full replacement is justified when a critical zone has not a single suitable angle or the recorder does not provide a stream via a standard protocol.

Reusing surveillance preserves cable routes, mounts, and power — installation does not shut down the area or require re-approvals, and launch time is reduced.

Engineers’ Concerns: False Triggers, Privacy, Load

Engineers usually have the same concerns: will recognition work in dust and on the night shift, where do recordings go and who sees them, how much manual work will false alarms leave, and will the existing network withstand a constant stream from cameras.

Below are answers to each point; they rely on the site’s engineering design — camera placement, zone marking, and retention policies — rather than on vendor promises.

Where are recordings stored and who has access to them?

We do not record or export the full stream: the archive stores event metadata — class, zone, time — and a short clip around it. Access is separated by roles: security sees the fact and zone, the occupational safety service sees the clip, and the manager sees a summary.

Clip viewing is written to a log, and the retention period is configured according to internal procedures and personal data requirements.

Will false alarms have to be handled manually?

No: in the interface, an event is marked as confirmed or false, labels accumulate and are used to retrain thresholds, so the share of false triggers falls from week to week.

Review takes minutes: the system filters out most disputed frames itself, leaving the engineer only genuinely ambiguous shifts.

Will the existing network have enough bandwidth?

Analysis is performed locally, on a node next to the cameras, so only events and short clips go to the shared network — a few percent of the video stream. Recording channels and the site’s operational traffic are not affected, and a separate communication line is not required.

Let’s Discuss the Task: We’ll Select a Scheme for Your Site

To make the conversation specific rather than general, it is enough to send what you already have. Describe the task — we will come back with an estimate and a work plan, without preliminary surveys or unnecessary approvals.

• layouts or zone diagrams where monitoring is needed: entrances, warehouses, areas with vehicles, high-risk zones;
• a list of cameras and recorders already installed, or requirements for new shooting points;
• which violations we detect first — absence of a hard hat, vest, safety glasses, entry into a restricted zone, person falling;
• how response is currently organized and who receives the notification: security post, dispatcher, shift manager;
• requirements for event storage and access: who sees recordings, whether a separate archive is needed, whether face blurring is required;
• site constraints: connectivity, power, lighting, dust, night shifts, outdoor work;
• indicators by which you will measure the result: share of detected violations, response time, number of walkthroughs.

Based on this set, we select a scheme: where to process frames — on site or on a server, how many camera points one machine can handle, how the system behaves during a communication outage, and what goes into the log.

Then we come back with an estimate and work plan: stage composition, what we configure at the first launch, when the first automatic detection appears, and how we expand to neighboring sites.

We start the project with a pilot zone — it shows recognition quality under real lighting and people flow.

If accuracy is sufficient under your conditions, we scale to the remaining areas with the same circuit; if not, we change camera points and settings before the solution goes across the entire site.