A context-aware safety layer deploys on existing CCTV in about an hour and can verify a defined set of height-work events: harness worn, anchor point used, edge zone entered. It cannot replace the permit holder who decides whether the work is safe to start. A spotter who is also the crew's deadline is not a control. That is the gap this layer fills, and the one it does not.
In Singapore and Malaysia, work at height is simultaneously the most-written safety rule and the most eyeball-supervised activity on site. The standard control is one spotter per crew, competing against the schedule they are meant to slow down. Supervision by presence does not scale, and supervision by camera raises coverage and privacy questions that EHS has to answer before a lens goes up, not after an incident.
What a camera can verify, and what needs a human
The useful distinction is not "AI watches the site". It is which specific events are machine-verifiable from a frame, and which require human judgment that a camera only records.
| Event | Camera can verify | Needs a human |
|---|---|---|
| Harness worn | Yes — harness present and donned on the worker in the zone | — |
| Anchor point used | Yes — lanyard connected to a designated anchor | Whether the anchor is rated for this load |
| Edge zone entry | Yes — worker crossed into a defined elevated-edge boundary | Whether a guardrail was required first |
| Ladder angle | Partially — gross angle and footing visible | Whether the ladder is the right equipment for the task |
| Permit validity | No | Permit holder confirms conditions, weather, exclusion below |
The honest reading of that table: cameras are good at "was the rule followed in view", poor at "was this safe to attempt". A VLM reads the scene and answers a specific question about it — is this worker's lanyard clipped to an anchor — which is why it catches states a fixed object-detector class list cannot enumerate. It still does not know if the anchor will hold.
The tension nobody states before buying
Camera verification of height work is a surveillance capability pointed at people, most of the time doing nothing wrong. Pretending otherwise is how these programs fail an audit or a works-council conversation. The resolution is governance decided before the lens is mounted:
- Scope — the layer verifies named events in defined zones, not continuous watching of individuals.
- Trigger — it activates on height-work activity in those zones, not as an always-on feed to a manager's phone.
- Retention — event evidence is kept to a stated window tied to the permit and the record requirement; raw footage is not an open archive.
- Access — who can see a live flag versus who can pull an after-the-fact record is written down first.
Decide those four and the cameras you already own produce a defensible evidence trail. Skip them and you have built a liability document.
Regulatory framing, named only
Singapore's Workplace Safety and Health (Work at Heights) Regulations govern fall-prevention planning, and Malaysia's Factories and Machinery Act 1967 with DOSH guidance covers the equivalent duties. This page names those instruments and invents no clause numbers. The point is not to quote a regulation back at you; it is that both regimes expect a documented, followed procedure, and an event record that shows harness, anchor and edge discipline is evidence that the procedure was followed, not a substitute for the permit system that sits above it.
Where this fits, honestly
Hypernology's AI Safety layer is built on a context-aware vision-language model and deploys on existing CCTV, which is why the setup is measured in an hour rather than an infrastructure project. The height-work event set above is reasoning about what is observable from a frame; the deployment proof Hypernology holds is in safety monitoring on existing cameras generally, not a published fall-from-height case. Treat the event checklist as the specification to test on your own site footage, not as a claimed result.
A spotter pulled onto the schedule is a person, not a control. A camera that witnesses the harness, the anchor and the edge is a control for exactly those events and nothing beyond them. Knowing the line between the two is the whole competence here.
Related reading: the lone-worker monitoring post, the machine-guarding compliance cluster, and /solutions/hyperq-ai-safety.
Frequently asked questions
Can a camera replace a work-at-height spotter? No. It verifies observable events — harness worn, anchor used, edge entered — and records them. A permit holder still judges whether the work is safe to start and whether the anchor is rated.
Is monitoring workers at height a privacy problem? It is a surveillance capability, so it requires a retention and access policy set before deployment: defined zones, event-triggered activation, a stated retention window, and named access. With those, it is defensible evidence; without them, it is a liability.
Does it work on our existing CCTV? The context-aware layer is designed to run on existing cameras, which is why deployment is measured in about an hour rather than a cabling project. Coverage still depends on camera placement and sightlines to the work zone.
Send one site's camera layout and the height-work zones you need covered, and within two weeks we will return which events your current sightlines can actually verify and the retention model to make them admissible, before any rollout. Start at https://apac.hypernology.net/contact.