AI video analytics can help a workplace see more, faster. It can flag a worker entering a restricted zone, a missing helmet, a vehicle-pedestrian conflict, smoke, a fall event, or repeated unsafe behaviour in a high-risk area.
But for Singapore workplaces, the key question is not only whether the system can detect something.
The real question is: what happens after the alert?
Detecting a helmet is easy. Knowing whether the work is actually controlled is the hard part.
Why This Matters
Singapore is already moving toward wider use of video-based workplace safety monitoring. MOM recognises video analytics as a workplace safety and health (WSH) technology category for detecting safety violations, hazards and incidents in real time. From 1 June 2024, construction worksites with contract value of $5 million and above must install video surveillance systems (VSS). MOM has also referred to SafeSite video analytics pilots and AI-supported hazard identification for construction safety.
These are important signals. Video-based monitoring is becoming part of the WSH operating environment, especially in visually monitorable sectors such as construction, logistics, manufacturing, marine, facilities management and industrial operations.
Still, technology does not replace statutory duties. Under Singapore's WSH framework, duty holders still need risk assessment, reasonably practicable controls, safe work procedures, competent supervision, incident response and corrective action. A camera or dashboard does not dilute those duties.
If an AI system repeatedly detects the same unsafe condition, that is not proof that the workplace is well controlled. It may be evidence that the risk is known and still not being fixed.
What Organisations Should Know
AI video analytics is best treated as an additional monitoring and escalation layer, not a replacement for risk management.
Every detection rule should be linked to a real hazard, risk assessment, safe work procedure or site rule.
False negatives are the bigger WSH danger because they create false reassurance.
False positives are the bigger adoption danger because they train supervisors to ignore alerts.
Alerts must have an owner, a response time, an escalation path and closure evidence.
High-severity alerts need clear stop-work or pause-work authority.
Footage, faces, body images, locations, badges and behavioural patterns may raise PDPA and worker-trust issues.
A simple test helps: if an alert has no defined response, it is not a safety alarm. It is only a notification.
Common Gaps We See
Buying detection before designing response. Many projects spend too much energy on what the model can detect, and not enough on who must act, how quickly, and with what authority.
Treating PPE detection as safety control. A worker wearing a helmet and vest may still be exposed to falling objects, moving vehicles, live energy, unsecured openings, heat stress, dust, poor housekeeping or suspended loads. PPE detection is useful, but it is not the same as risk control.
Ignoring camera and site limitations. Lighting, rain, glare, dust, occlusion, camera angle, blind spots, frame rate, site changes and model drift can all affect performance. No alerts does not always mean no unsafe work.
Creating alert fatigue. If supervisors receive too many nuisance alerts, they may start muting the dashboard, acknowledging without checking, or reviewing clips only after the shift. A noisy system can reduce trust in real alerts.
Leaving privacy until later. Workplace video analytics should not be treated as just an IT tool. Organisations need clear purposes, worker notification, access control, retention rules, vendor controls, export logs and appropriate limits on how footage is used.
Practical Steps To Consider
Start with the risk register, not the camera catalogue.
Identify the hazards worth monitoring: lifting zones, vehicle-pedestrian interfaces, work-at-height access points, restricted hazardous areas, fire or smoke risk areas, loading zones, or repeated unsafe locations.
Create a rule register.
For each detection rule, define the hazard, linked risk assessment or safe work procedure, camera or zone, severity, alert owner, backup owner, response time, required action, evidence and review frequency.
Define stop-work thresholds.
Some alerts should not wait for a slow dashboard review, such as a person inside an active lifting exclusion zone, worker near an unprotected edge, smoke/fire detection, worker collapse, or entry into a restricted hazardous area.
Pilot under real conditions.
Test day and night, rain and glare, peak activity, partial occlusion, different PPE types, different work phases and actual supervisor workload. Track true positives, false positives, missed events, response time, closure quality and worker feedback.
Design privacy and trust controls early.
Explain the safety purpose, notify affected workers and visitors, restrict raw footage access, define retention, control vendor processing, log exports, and provide a way to challenge wrong alerts.
Review trends, not just incidents.
Repeated alerts should trigger risk assessment review, contractor coordination, toolbox learning or management intervention. The goal is reduced risk, not a busier dashboard.
How DASH Consult Can Help
DASH Consult supports organisations that want to use safety technology without losing sight of the actual WSH control system.
We can help review whether AI video analytics rules are hazard-led, linked to risk assessments, supported by supervisor response procedures, aligned with stop-work authority, and governed with appropriate privacy and evidence-handling controls.
The useful question is not "Can the AI detect it?"
The better question is: does the alert lead to timely action that reduces risk and can withstand review?
FAQ
Is AI video analytics mandatory for all Singapore workplaces?
No. MOM recognises video analytics as a WSH technology category, and VSS requirements apply to certain construction worksites, but organisations should avoid implying that AI video analytics is mandatory for every workplace.
Can AI video analytics replace supervisors?
No. It can support monitoring and escalation, but competent supervision, risk assessment, safe work procedures, engineering controls and stop-work authority remain essential.
Are false positives a serious issue?
Yes. False positives can overload supervisors, frustrate workers and train people to ignore the system. A safety alert channel must stay useful and credible.
Does a safety purpose remove PDPA concerns?
No. Safety may support the purpose of monitoring, but organisations still need to manage notification, purpose limitation, protection, retention, access, vendor controls and transfer issues where personal data is involved.
What should a good deployment include?
At minimum: a hazard-led use-case register, camera coverage map, detection performance review, alert response SOP, stop-work matrix, supervisor training, worker communication, data protection controls and a post-deployment review schedule.