Construction site safety monitoring

• Cameras on plant and at fixed points watch the agreed high-risk zones: plant exclusion zones and haul routes, open edges, designated hard-hat and hi-vis zones, and lift zones during lifts.
• Detection runs on site, on an edge box in the cabin and a small unit on each machine, so alerts arrive in seconds with no dependence on the site's internet link.
• When a person is detected inside a plant's exclusion zone, the operator gets an in-cab alert and the zone's supervisor gets a notification with a short clip. They decide what to do.
• A broken guard-rail line at an open edge, or people inside a lift zone while a lift is in progress, is flagged to the supervisor to check and act on, and logged as an observation in the health and safety system.
• Detections and near-miss reports are combined into trends by zone, activity, phase and time of day, so the weekly look-ahead meeting can plan out the hot spots before the next phase starts.
• The system never identifies anyone. No facial recognition, no matching to names or badges, faces blurred before anything is stored, and nothing is used to score or discipline individual workers.

• HSEHealth and Safety at Work etc. Act 1974 and CDM Regulations 2015 (regulations 12 to 15)
  →

  The principal contractor must plan, manage, monitor and coordinate the construction phase, and record the arrangements in the construction phase plan. Camera monitoring is one control in that plan, not a substitute for supervision, inspections, permits or segregation. HSE's statement on its approach to AI (June 2026) is clear that health and safety law applies to AI, so the system and its failure modes get their own risk assessment. Regulation 14 requires the principal contractor to consult and engage with workers, which is where this system is introduced.

• HSECDM 2015 regulations 27 and 28 (traffic routes and vehicles) and PUWER 1998
  →

  People and plant must be kept apart by the site layout first. Cameras spot when that separation fails; they don't replace it. Fitting cameras and alert units to plant changes work equipment, so do it with the plant owner, keep the operator's view clear, and never connect the unit to the machine's controls. Any engineered proximity system that already slows or stops a machine stays in place and is outside this design.

• HSEWork at Height Regulations 2005 and the PPE at Work Regulations 1992
  →

  The order is: avoid work at height, prevent falls with collective protection such as guard-rails, then minimise the consequences. A detected gap in edge protection prompts a supervisor to check and fix it; it never counts as the inspection the regulations require. PPE is the last line of protection, so PPE alerts are limited to designated zones and treated as a sign that a higher control has failed.

• HSELOLER 1998 (regulation 8: organisation of lifting operations)
  →

  Every lift must be planned by a competent person, supervised and carried out safely. The lift plan defines the exclusion zone; the system checks that zone during the lift and alerts the lift supervisor and the crane operator. They decide whether to stop. The system never signals a crane or tells a slinger-signaller what to do.

• ICOUK GDPR, the ICO's monitoring workers guidance and video surveillance guidance
  →

  Camera images of people are personal data, and many of the workers belong to subcontractors. That means a DPIA before any camera goes live, a clear purpose written down, the least intrusive means, signs and induction briefings so everyone on site knows, no covert monitoring, short retention, and views sought from workers and their representatives. The ICO says its monitoring guidance is under review following the Data (Use and Access) Act 2025.

• ICOBiometric recognition guidance (February 2024)
  →

  Using facial recognition, or any biometric feature, to identify people is processing special category biometric data, which needs an additional legal condition and is hard to justify on a site. This design avoids it entirely: the models detect a person, a hard hat or a hi-vis vest, never who someone is. Faces are blurred on the edge box before any clip is stored.

• BISTEmployment Rights Act 2025 and the workplace monitoring consultation
  →

  The Act received Royal Assent on 18 December 2025. It sets no specific rule on site cameras, but the government consulted from July to September 2026 on whether employers should have to consult representatives before introducing monitoring technology. Consulting safety and worker representatives from discovery, and agreeing in writing that footage is not used for discipline, fits any likely outcome.

On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →

• A written purpose statement, agreed with safety and worker representatives before the pilot: what cameras watch, what they don't, and that footage is not used for discipline or to score anyone.
• Site inductions, signs at every entrance and the privacy notice explain the cameras to everyone, including subcontractors' workers and visitors. Subcontract terms say the same.
• A response rule for each alert type: who gets it, what they check, and how they record what they did.
• A named owner for the detection zones on each site, who updates them when the layout changes, as part of the daily or weekly planning routine.
• A weekly review of false alarms and missed events by zone and hazard type, with the health and safety manager owning each model and its thresholds.
• Camera checks added to the daily site inspection: lens clean, still pointing at its zone, still covered by the plan.

1. 3–4 weeks

   Discovery

   Pick one site and two hazards (usually plant exclusion zones and lift zones), walk the site with supervisors and operators, test camera views in real conditions, measure today's near-miss reporting, and brief safety representatives.

2. Gate · must be true to passHazards and zones agreed, DPIA started, purpose statement drafted with representatives, plant owners agree the in-cab fit.

3. 10–12 weeks

   Pilot

   Cameras and edge detection on one site in shadow mode first: alerts go only to the health and safety team, who compare them with what really happened. Then live alerts to operators and supervisors for the two hazards.

4. Gate · must be true to passAlert precision and time to act within agreed limits, no use beyond the agreed purpose, representatives' feedback heard, DPIA signed off.

5. 4–6 weeks

   Production

   Live on the pilot site with monitoring, support, the weekly review pack and a rollback plan. Near-miss trends added once there's enough history.

6. Gate · must be true to passSupport model, zone ownership and supervisor training in place; rollback tested.

7. Ongoing

   Scale

   Add open-edge and PPE-zone detection, then more sites, reusing the edge build and the purpose statement.

8. Gate · must be true to passEach new hazard and site passes the same pilot gate, including consultation.

R ResponsibleA AccountableC ConsultedI Informed· not involved

Where I'd fit

1. 1

   Goes wrong

   Workers hear about the cameras on the day they appear, assume they're being watched for discipline, and stop reporting near misses.

   How the design avoids it

   Consult representatives from discovery, publish the purpose statement, and stick to it. Never use the footage for discipline.

2. 2

   Goes wrong

   Too many false alarms, so operators mute the in-cab unit and supervisors ignore notifications.

   How the design avoids it

   Shadow mode first, two hazards only, and a weekly review of false alarms with the people who receive them.

3. 3

   Goes wrong

   The site changes every week, but the zones don't, so the cameras watch yesterday's layout.

   How the design avoids it

   A named owner for zones on each site, zone changes in the planning routine, and coverage measured daily.

4. 4

   Goes wrong

   The demo used clear daylight footage. The site has rain, dust, glare and dark winter afternoons.

   How the design avoids it

   Prove each hazard can be seen in footage from your own sites, in winter, before choosing a model.

5. 5

   Goes wrong

   The system quietly becomes the control: supervision thins out because the cameras are watching.

   How the design avoids it

   Record it in the construction phase plan as an extra layer, and keep supervision, inspections and permits exactly as they were.

1. From 10 November 2026
   Azure IoT Edge 1.5 LTS reaches end of support. IoT Edge 1.6 LTS is supported to 14 November 2028.

   If you run detection on IoT Edge, build on 1.6 LTS or Azure IoT Operations, and plan an upgrade path for every edge box, since site kit moves from job to job.

2. Response awaited
   The government's consultation on workplace monitoring technologies (8 July to 30 September 2026) asked whether employers should have to consult staff representatives before introducing monitoring tools, and whether guidance, a statutory code or new law is needed.

   Consult safety and worker representatives from discovery and keep the purpose statement and DPIA up to date. That fits any likely outcome.

3. Likely from 2028
   HSE consulted (7 April to 30 June 2026) on reforming RIDDOR, including new reportable dangerous occurrences for construction such as plant overturning and objects falling from structures under construction. A response is awaited, and changes need new regulations.

   Use event and near-miss categories that can map to new reporting types, and keep RIDDOR decisions with the responsible person, not the system.

4. 2026–2027
   The ICO's monitoring workers guidance is under review after the Data (Use and Access) Act 2025, and the ICO must produce a statutory code on AI and automated decision-making. HSE's Regulation Action Plan (July 2026) commits to clarifying its expectations on AI.

   Keep the DPIA, purpose statement, access logs and model validation records in a form you could show the ICO, HSE or a safety representative.

5. 25 September 2028
   Azure Custom Vision and the Azure Vision Image Analysis API (v3.2 and v4.0) retire.

   Don't build on them. Train models in Azure Machine Learning, SageMaker AI or an open-source stack, in a portable format such as ONNX.