DEVB TC(W) No. 3/2023, Annex A 5(12)(g)
AI cameras: the Safety Monitoring System using Artificial Intelligence
Clause 5(12)(g) of DEVB TC(W) No. 3/2023, Annex A, requires IP cameras with AI processors to detect unsafe scenarios on site in real time, such as people in danger zones, missing PPE and work at height without a harness. Accuracy must pass 80% in testing and average 90% after four weeks. Kodifly supplies, installs and tunes them.
What the circular requires, and how sites use it
The contractor installs enough IP cameras, with AI processors, protective casings and sensor systems, to cover all areas of the site for real-time monitoring. A minimum camera number is set in the contract (clause 5(12)(g)(i)). Cameras have a resolution of at least 1080 pixels, at least 5x optical zoom and encrypted live streaming to the platform (clause 5(3)).
On a Hong Kong site, the cameras watch gates, plant routes, lifting areas and work at height. When the AI sees an unsafe scenario, a siren and red light go off at the spot, supervisors get an alert, and the platform keeps the clip. The Buildings Department and CIC name this safety monitoring system using artificial intelligence.
Unsafe scenarios in Annex A, clause 5(12)(g)(ix)
The clause reads including but not limited to. The AI must detect several events at the same time.
| Scenario | What the AI watches |
|---|---|
| Zone entry | Unauthorised access to restricted zones, danger zones, lifting zones and no-parking zones. Zones and alerts are configurable. |
| Workers near plant | Workers near site vehicles or plant designated by the contractor, and potential collisions between workers and vehicles or plant. |
| Driver and operator state | Fatigue, distraction and inattentive behaviour of site vehicle drivers and plant operators. |
| PPE | Workers and others without the required safety helmet and safety vest. |
| Lifting height | Lifting heights above the authorised limits. |
| Work at height | Work at height without a safety harness fixed to a proper anchorage point, or without a proper ladder platform or working platform. |
| Smoking | Smoking within the construction site. |
| Fire and smoke | Fire and abnormal smoke in specified areas. Used only when the smart fire and smoke detection system is not in the contract. |
The accuracy ladder
The circular sets tests and a long tuning period.
Before go-live
Test
At least 80% true positive for each unsafe scenario, on a dataset of at least 10,000 images, half with and half without the object or event (clause (iii)).
Within 2 weeks
Train
Two on-site technical courses of 4 hours each (clause (viii)).
From week 4
Hold
At least 90% average accuracy across all required scenarios, kept to completion (clause (vii)).
Weekly, then bi-weekly, then monthly
Improve
An AI improvement plan with re-training goes to the Project Manager or equivalent (clauses (iv) to (vi)).
Every 6 months
Drill
A drill tests whether the AI safety monitoring works (clause (xvi)).
How Kodifly supplies, installs and runs it
Cameras, AI processors and the platform are one project. Alarms and records follow the clause.
Devices, software and installation together
IP cameras and AI processors with protective casings
Zones drawn on screen or by plastic barriers
Siren and red light within 1 second
Alerts to supervisors and the Safety Officer
Encrypted video from 1 minute before to 1 minute after
Response time record and summary report
Accuracy test, tuning and training
AI cameras, answered
What does the circular require of AI safety monitoring?
IP cameras with AI processors covering all areas of the site, detecting unsafe scenarios in real time, with a siren and alerts, encrypted video and a response record (Annex A, clause 5(12)(g)).
What accuracy must the AI reach?
At least 80% true positive for each unsafe scenario in the live test, on a dataset of at least 10,000 images. From week 4, the average across all required scenarios must be at least 90% (clauses (iii) and (vii)).
How fast must an alarm sound?
The sound and light signals trigger within 1 second of detecting an unsafe scenario, with a siren of at least 70 dB, or a higher level agreed, and a flashing red light (clause (xii)).
How is video kept?
Recording of an unsafe scenario runs from 1 minute before to 1 minute after. The video is stored, live-streamed to the platform and encrypted (clauses (xiii) and (xiv)).
How are danger zones defined?
On the screen, by the user, or by the system recognising a series of plastic barriers placed on site (clause (xi)).
What must the cameras themselves meet?
A protective casing, resolution of at least 1080 pixels, at least 5x optical zoom, and continuous wireless encrypted live streaming to the platform (clause 5(3)).
Does the AI also detect fire and smoke?
Only where the contract has no separate smart fire and smoke detection system. Otherwise the AI cameras can be an integrated part of that system (clauses (ix) and 5(12)(k)(ii)).
Related guides
Plan your AI camera cover
Send us the site layout and the unsafe scenarios in your contract. We will plan the cameras and the test.