1 month. That is how long it took a Busan-based manufacturer to go from signed contract to live confined-space monitoring — on cameras already mounted in the facility. No new hardware procurement, no new camera vendor, no additional civil works. The software auto-recognised every ONVIF-compliant camera on the network and started watching entry events, intrusion, and man-down conditions from day one.
For safety managers operating under Singapore's Workplace Safety and Health (WSH) framework, that timeline matters. Confined-space entries are the highest-scrutiny activity on any industrial site. The entry-permit system, the standby-person requirements, and the atmospheric testing obligations create paperwork and accountability chains that manual processes handle poorly at scale. Most sites already have CCTV. The gap is not cameras — it is software that turns passive footage into active monitoring.
This post explains how AI-based confined-space monitoring works on existing CCTV infrastructure, what WSH obligations it addresses directly, and what it cannot replace. It also includes a practical monitoring checklist mapped to the key WSH duty categories, so safety officers can evaluate coverage before selecting a solution.
Why confined-space entries attract the most scrutiny under Singapore WSH
Singapore's WSH Act and its subsidiary regulations treat confined-space work as a distinct high-risk category. The relevant obligations cluster around three roles: the responsible person (RP) who authorises and supervises the work, the competent person who performs atmospheric testing and signs off on the permit, and the standby person who monitors from outside the confined space during entry.
The standby person duty is where video monitoring becomes operationally relevant. A standby person must maintain continuous visual or voice contact with the person inside, be ready to initiate rescue, and not enter the confined space themselves during the watch period. On large sites with multiple simultaneous entries — vessel inspections, manhole work, tank cleaning — maintaining qualified standby personnel at every point strains headcount. CCTV covers the visual-contact requirement electronically, but only if the feed is actively watched and can trigger an alarm automatically.
SS 580, the Singapore standard on confined-space safety, reinforces the requirement for continuous monitoring and documented response procedures. The permit-to-work system requires time-stamped entry and exit records. Both create compliance documentation burdens that manual logbooks address inconsistently — especially on night shifts or during high-volume maintenance windows.
The gap that AI monitoring fills: it watches continuously, logs events with timestamps automatically, and raises an alert the moment a man-down event, unauthorised intrusion, or missed exit is detected. It does not replace the standby person's judgment or rescue capability. It makes the standby person more effective by removing the cognitive load of continuous visual scan.
What ONVIF auto-recognition means in practice
ONVIF is the open-standard protocol that most industrial and commercial IP cameras have supported for over a decade. When a monitoring platform claims ONVIF auto-recognition, it means the software connects to the camera network, discovers compatible devices automatically, and begins pulling video streams without requiring per-camera configuration.
HyperQ AI Safety discovers ONVIF cameras on the local network during initial setup. No camera-by-camera credential entry, no vendor-specific SDK for each camera brand. For a facility with 30 cameras across a mix of manufacturers — which is the normal situation after five or more years of incremental CCTV expansion — this matters. The setup time is measured in hours, not weeks.
The Busan-based manufacturer's 1-month deployment reflects that reality. The site had existing CCTV infrastructure covering production areas and confined-space entry points. The implementation timeline was driven by configuration, user training, and alarm-routing integration, not by camera installation or civil works. Software pricing starts from $10,000 — compared with a new hardware-locked safety-camera stack that would run $20,000 or more before installation.
The detection classes active in confined-space contexts are: fall detection, man-down (motionless person), fire and smoke, PPE detection (helmet, vest, harness), and zone intrusion. Each is model-based, not rule-based. A rule-based system flags on pixel-change thresholds; it generates false alarms from shadows, lighting changes, and camera vibration. A model-based system recognises body posture, object shape, and thermal signature. False-positive reduction is 60–80% compared with threshold-based alternatives — which determines whether operators learn to trust the alerts or learn to ignore them.
The real cost comparison: software versus new hardware
A site safety manager making the case for confined-space monitoring investment has two paths.
Path 1: New dedicated safety cameras. A purpose-built safety-monitoring system from a hardware-locked vendor involves camera hardware ($420–$2,250 per unit depending on type, plus thermal at $2,250 and IR at $650 per unit), mounting, cabling, network provisioning, and the vendor's software licence. On a mid-size industrial site with 10–15 confined-space monitoring points, total spend exceeds $20,000 before any professional services. Lead time from purchase order to live monitoring is typically 3–6 months.
Path 2: AI software on existing CCTV. HyperQ AI Safety software starts from $10,000. If the existing cameras are ONVIF-compliant (the majority of commercial IP cameras installed after 2012 are), no additional hardware is required for standard visual detection. Smartband biometric monitors — which add heart rate, blood oxygen saturation (SpO2), and skin temperature for personnel inside confined spaces — are available from $35 to $250 per unit. The Busan deployment was live in 1 month.
The hardware cost saving on Path 2 versus Path 1 ranges from 30–50% depending on how many existing cameras are reusable. ROI payback on the software investment runs 11–18 months on sites where it replaces or augments a manual standby-person roster.
The honest caveat: if existing cameras are positioned incorrectly for confined-space entry-point coverage, some repositioning or supplemental cameras may still be needed. ONVIF auto-recognition solves the software integration problem — it does not solve a physical coverage gap. Before any procurement decision, a camera-coverage audit against confined-space locations is the right first step.
Confined-space monitoring checklist mapped to WSH duties
The table below maps key WSH duty categories to the monitoring capabilities that address them. It is designed as an evaluation tool, not a compliance sign-off — confirm specific regulatory obligations with a qualified WSH officer.
| WSH duty category | Monitoring capability needed | HyperQ AI Safety coverage |
|---|---|---|
| Continuous visual contact with entrant | Real-time video monitoring of entry point and interior | Yes — camera-based, continuous, logged |
| Man-down / incapacitation detection | Automated alert on motionless-person detection | Yes — model-based posture recognition |
| Unauthorised entry prevention | Zone-intrusion alert before entry without permit | Yes — configurable entry-zone perimeter |
| PPE compliance at entry point | Helmet, vest, harness detection at gate/access point | Yes — PPE class detection on entry frames |
| Atmospheric hazard early warning (fire/smoke) | Smoke and fire detection inside confined space | Yes — visual smoke/flame detection |
| Biometric monitoring inside confined space | Physiological alert on heart rate / SpO2 anomaly | Yes — smartband integration, $35–250/unit |
| Time-stamped entry/exit record | Automated event log with timestamps | Yes — logged to dashboard, exportable |
| Permit-to-work time compliance | Alert on entry outside permitted time window | Yes — configurable time-window rules |
| Night-shift and low-light coverage | IR or thermal camera support | Yes — thermal ($2,250) and IR ($650) per unit |
| Multi-site / multi-entry-point coordination | Centralised dashboard, multi-camera view | Yes — web dashboard, no per-site server |
Two duties are outside software scope and remain human obligations regardless of monitoring technology:
- Atmospheric testing: the competent person must physically test oxygen levels, flammable gas, and toxic gas concentrations before each entry. No camera-based system substitutes for gas detection instruments.
- Rescue capability: the standby person must be physically present, able to initiate rescue without entering the space, and trained in rescue procedures. Monitoring software triggers the alert — it does not execute the rescue.
These limitations are not edge cases. Any vendor who implies software alone satisfies the competent-person or standby-person obligation under Singapore WSH is misrepresenting the regulatory requirement.
How the detection works: architecture without the marketing
HyperQ AI Safety runs on a server or edge device on the local network. Video streams are pulled from ONVIF cameras in real time. Detection models run inference on each stream — not on stored footage, on the live feed — so alert latency is sub-second on standard hardware.
Each detection class is a separate model. Fall detection reads body-pose keypoints and classifies posture transitions from upright to horizontal. Man-down detection identifies a person who has been stationary on the ground beyond a configurable time threshold. Zone intrusion detection uses a drawn perimeter on the camera view; any person entering the zone outside a permitted time window triggers an alert. PPE detection runs an object-detection model on frames at the entry checkpoint — it identifies presence or absence of helmet, vest, and harness classes.
Smartband integration adds a data layer that cameras cannot provide. A person wearing a smartband transmits heart rate, SpO2, and skin temperature to the platform continuously. Physiological alert thresholds — for example, SpO2 below 94%, heart rate above 130 bpm — are configurable per confined-space context. The smartband channel is the only way to detect physiological distress before it becomes a man-down event visible on camera.
The combination is what makes confined-space monitoring credible: camera-based detection for external events (entry, intrusion, fire, PPE) and wearable biometrics for internal physiological state. Either alone leaves a gap.
What the Busan deployment actually looked like
A Busan-based manufacturer — operating in a sector where confined-space access for equipment maintenance is routine — deployed HyperQ AI Safety on its existing CCTV infrastructure. The implementation ran from signed contract to live monitoring in 1 month.
The implementation steps: network discovery of ONVIF cameras, zone configuration for each confined-space entry point, alert routing to the safety officer dashboard and mobile notifications, and operator training. No new cameras were installed. The site's existing IP camera network, accumulated over several years from multiple vendors, was fully compatible through ONVIF auto-recognition.
The deployment did not eliminate the standby person role. It augmented it. The standby person's attention could shift from passive visual monitoring of a screen to active response readiness, because the software was handling the continuous scan and would push an alert to their device the moment an anomaly was detected.
For Singapore WSH compliance context, sites in similar industrial sectors operating under comparable regulatory regimes have found the same architecture applicable. The construction safety compliance context in neighbouring jurisdictions reinforces the pattern: regulators are moving toward documented, technology-backed monitoring rather than relying on manual logbooks for high-risk work categories.
Where this approach does not work
Three scenarios where reusing existing CCTV for confined-space monitoring runs into hard limits:
Camera positioning is wrong. Existing cameras installed for perimeter security or general floor surveillance may not cover confined-space entry points at the angles needed for body-pose detection. Fall detection requires a camera with a view of the full body — a ceiling-mounted camera directly above a manhole works; a camera 30 metres away at an oblique angle does not. A coverage audit must come before a software decision.
Existing cameras are below minimum resolution. Body-pose detection models require sufficient pixel density over the monitored zone. Older cameras running at 720p or below, placed at distance, may not resolve body keypoints reliably. The software will flag low-confidence detections — this is not a failure mode that gets hidden, but it does mean some camera upgrades may be necessary.
Very deep or enclosed confined spaces with no camera line-of-sight. Tanks, vessels, and deep pits where a camera cannot see the interior require either a camera lowered into the space (which introduces its own installation complexity) or reliance on smartband biometrics alone for internal monitoring. Neither is an edge case on refinery, wastewater, or vessel-maintenance sites.
Honest evaluation of these three points before procurement saves a site from deploying software and finding the coverage gaps after go-live.
Frequently asked questions
Does AI monitoring on CCTV satisfy the standby-person requirement under Singapore WSH?
No. The standby person must be physically present, trained in the emergency response procedure, and capable of initiating rescue. Camera monitoring augments the standby person's ability to detect events — it does not substitute for the role. Regulatory obligation remains with the appointed competent person and standby person.
What does ONVIF auto-recognition actually require on the network?
The cameras must be ONVIF-compliant (most commercial IP cameras manufactured after 2010 are) and accessible on the same network segment as the HyperQ AI Safety server or edge device. Standard ONVIF Profile S or Profile T is sufficient for live-stream access. Camera credentials are required during initial discovery.
Can the system run without internet connectivity?
Yes. HyperQ AI Safety runs on a local server or edge device. Detection inference is on-premises. Cloud connectivity is optional for remote dashboard access and software updates — it is not required for real-time detection and alerting.
How long does initial configuration take for a site with 20 existing cameras?
Based on the Busan deployment and similar implementations, network discovery and initial zone configuration for 20 cameras runs 1–2 days on-site. Alert routing, dashboard setup, and operator training add another 1–2 days. The 1-month deployment figure includes contract finalisation, travel scheduling, and user acceptance testing — active on-site setup is 2 days.
What happens if a camera goes offline mid-shift?
The platform monitors camera-stream health and raises an alert on camera loss. The standby person and safety officer are notified immediately. The monitoring gap is logged. This is a material improvement over a manual monitoring regime where camera downtime may go unnoticed for an entire shift.
Does smartband data integrate with the permit-to-work system?
Currently smartband data flows into the HyperQ AI Safety dashboard and can be exported. Direct API integration with third-party permit-to-work platforms is a configuration project — contact the team to assess compatibility with your existing EAMS or PTW system.
The decision before buying anything
Before a safety manager on a Singapore WSH site evaluates any monitoring product, two questions settle most of the procurement decision:
- Are your confined-space entry points within field-of-view of existing ONVIF cameras at sufficient resolution and angle for body-pose detection? If yes, software alone closes the monitoring gap.
- Do the people entering confined spaces need physiological monitoring (deep spaces, oxygen-depleted atmospheres, heat stress risk)? If yes, add smartbands at $35–250 per unit to the software deployment.
If both answers are yes, the total investment is software from $10,000 plus smartbands. No new camera procurement, no new vendor, no civil works. The AI safety monitoring on existing CCTV (HyperQ AI Safety) page has the technical specification, detection class list, and ONVIF compatibility details.
If the camera-coverage audit reveals gaps, the marginal cost of filling them with ONVIF-compatible cameras is significantly lower than replacing the entire monitoring infrastructure with a hardware-locked system.
The entry log filled in by hand every shift is a compliance artefact. The CCTV watching the manhole can generate that log automatically and raise the alarm before a man-down event becomes a fatality. The cameras are already there. The software question is what to do with them.
Send us your confined-space entry-point camera layout and a floor plan of the monitored zones. We will map detection coverage, identify any angle or resolution gaps, and return a deployment assessment within 5 business days — no contract required until the coverage spec is confirmed against your actual site. Start the assessment at apac.hypernology.net/contact.