HyperQ AI Safety deploys to a production floor in 1 hour. The harder problem is that most manufacturing facilities in Singapore and Malaysia have no automated near-miss detection at all -- not because the technology is expensive, but because the process of identifying near-misses still depends on a worker who witnessed an event choosing, on their own time, to file paperwork that benefits neither them nor their shift team in any measurable way that day.
Near-miss logs at most sites read clean. A forklift driver brakes hard to avoid a pedestrian crossing the aisle, both workers exchange a look, and both get back to work. No record, no analysis, no corrective action. The near-miss count in the monthly safety dashboard reads zero -- not a safe floor, just an unreported one.
This post covers what near-miss detection as a camera capability actually means: the event classes a system classifies reliably, how those map to Singapore's WSH Reporting of Dangerous Occurrences framework and Malaysia's DOSH requirements, and how a continuous feed of classified near-miss events becomes the leading-indicator data your injury rate can never produce.
Why the near-miss log stays empty
Workers are rational, not negligent. Filing a near-miss report requires stopping work, locating the correct form, writing a description accurate enough to be useful in a later review, and submitting it through a chain that may or may not acknowledge receipt. There is no compensation for this. In a high-output environment, the social cost is real: filing a report on a colleague's behavior can create friction that outlasts the incident.
A production safety manager described the resulting dynamic plainly: a shift team that files zero near-miss reports in a month does not attract scrutiny the way a shift team that files twelve would. The reporting infrastructure, without anyone designing it this way, rewards silence.
The consequence is significant. Near-miss frequency is your highest-value leading safety indicator. Every serious injury is typically preceded by a cluster of near-misses involving the same hazard, location, or workflow gap. If those near-misses are not captured, the pattern is invisible until someone is hurt.
Your injury rate tells you what already happened. Near-miss frequency tells you what the floor is producing right now. If you do not collect it, you cannot act on it.
Camera-based near-miss detection removes the human reporting step. The system observes continuously. Events are classified against defined event classes and logged with a timestamped record and the relevant camera feed, regardless of whether anyone present chose to report.
Near-miss detection as a camera capability: event classes
Near-miss detection via video analytics is an event-classification problem. The system observes position, velocity, trajectory, machine state, and zone occupancy. It does not observe intent. An event fires when the observable signals match a near-miss signature for a defined class.
The six primary event classes for manufacturing floor deployments, with their observable signals, regulatory mapping, and the HyperQ AI Safety detection mechanism, are shown below.
| Event class | Observable signals | WSH / DOSH regulatory category | HyperQ AI Safety mechanism |
|---|---|---|---|
| Forklift-pedestrian proximity | Zone breach; speed delta; trajectory convergence | MOM WSH Fourth Schedule: collision between powered vehicle and person | Vision zone monitoring + velocity estimation across overlapping camera views |
| Dropped load | Object trajectory; descent rate; load-zone occupancy | MOM WSH Fourth Schedule: fall of heavy article onto person | Motion trajectory classification; load-area zone mapping |
| Bypassed machine interlock | Machine-state signal + personnel entry while machine active | MOM WSH dangerous occurrence: machinery guard circumvented | Machine-state correlation + proximity event fusion |
| Unauthorized restricted-area entry | Personnel position inside exclusion zone without permit | MOM WSH / DOSH prohibited-area violation | Access zone monitoring with personnel classification |
| Slip/trip precursor | Gait anomaly; floor-state overlay (spill or wet surface) | General WSH risk: floor-level hazard | Behavioral pattern classification; floor-condition overlay |
| Vehicle-vehicle near-collision | Speed + trajectory convergence; gap below calculated stopping distance | DOSH factory hazard: multi-vehicle traffic management | Multi-camera triangulation; gap-to-stop estimation |
Each event produces a record: timestamp, event class, zone coordinates, classification confidence score, and the video clip covering the event window. That record is the input to investigation and corrective action. It is not a replacement for human judgment -- it is a reliable prompt for it.
How HyperQ AI Safety detects and alerts
HyperQ AI Safety runs on commodity camera hardware priced at $420, $1,200, or $2,250 per unit depending on resolution, housing requirement, and environmental specification. Edge inference runs on a dedicated processing unit inside your facility. Video does not route through a cloud service for real-time event classification; classification happens on-site.
Deployment to first event classification is 1 hour for a standard single-zone layout. That covers zone configuration, event-class calibration, and detection threshold setting for a defined floor area. Multi-zone deployments with complex vehicle routing or multiple exclusion areas require more time; there is no fixed multiplier because zone complexity varies considerably between sites.
The system does not require a historical dataset of near-miss footage for initial calibration. Event-class models transfer from training across a range of industrial environments and are tuned to your zone layout at deployment.
For immediate supervisor response, HyperQ AI Safety integrates with the $250 smartband worn by floor workers and supervisors. When a proximity or dangerous-occurrence event fires, the smartband generates a tactile alert and the supervisor dashboard updates in real time. For forklift-pedestrian proximity events, the reaction window can be under three seconds -- a smartband alert to a supervisor arriving in under one second is operationally relevant in a way a daily log review is not.
What the system does not do: it does not make enforcement decisions, it does not identify individual workers by name in its standard configuration (PDPA and PDPD compliance settings govern this), and it does not replace the investigation process that follows a flagged event. It creates the record and preserves the evidence.
Singapore WSH: reporting of dangerous occurrences
The Workplace Safety and Health Act requires employers to notify MOM of dangerous occurrences and workplace accidents. The WSH (Incident Reporting) Regulations set out the reporting scope; the Fourth Schedule defines the categories that constitute dangerous occurrences, including equipment collapse, pressure-equipment failure, and any occurrence in which a person could have been killed or seriously injured.
Several near-miss event classes sit directly adjacent to Fourth Schedule categories. A forklift-pedestrian near-collision that produces no injury is not a notifiable dangerous occurrence under current WSH rules. A forklift-pedestrian collision that injures the worker is. The line between those two outcomes is frequently a question of timing: did the driver brake in time?
MOM WSH's long-term strategy -- the WSH 2028 roadmap -- calls explicitly for industry to move from lagging indicators (injury rates, lost-time frequencies) toward leading indicators (near-miss frequency, hazard observation rates). The roadmap identifies near-miss reporting as a capability gap across manufacturing SMEs in particular. Automated near-miss detection is the operational mechanism that closes that gap without relying on voluntary worker reporting.
When a WSH enforcement investigation is triggered by an actual incident, the practical value of an existing near-miss log is immediately apparent. The investigator can review the frequency of forklift-proximity events in the relevant aisle before the injury occurred, whether that frequency was rising, and whether corrective action was documented after prior detections. Without automated detection, that history does not exist. The post-incident review operates on reconstructed accounts.
The video record produced by HyperQ AI Safety for each classified event is timestamped and preserves the full event window. Its usefulness in a formal MOM investigation depends on the specific circumstances and your legal counsel's guidance; the system is not designed as a legal evidence system, but the structured record it produces is substantially more useful than no record.
Malaysia DOSH: the OSH Act 1994 framework
Under Malaysia's Occupational Safety and Health Act 1994, employers must report occupational accidents resulting in death or four or more days of incapacity to DOSH. The Factories and Machinery Act adds requirements for specific equipment categories. Malaysia does not currently mandate near-miss reporting, but DOSH's own industry guidance and the National OSH Council's strategic plans consistently identify near-miss data as the primary target metric for proactive safety management -- noting that injury statistics arrive too late to prevent the next incident.
For manufacturers operating across both Singapore and Malaysia -- a common structure for regional suppliers with production in both countries -- a single automated detection system produces consistent, structured event records regardless of jurisdiction. The event-class definitions and log format are the same in both environments. Country-specific regulatory reporting is handled separately by your safety team; the detection system feeds both environments from the same underlying data.
Building a leading-indicator program from detected events
A near-miss detection system produces data. That data improves floor safety only when someone uses it to find patterns and act on them. Three analysis layers are worth establishing from the start.
Frequency trending by event class and zone. If forklift-proximity events in Aisle 4 increase from three in a week to eleven the following week, the floor has changed. Something is different about how vehicles or people are moving through that aisle -- a new delivery schedule, a layout change, a staffing rotation. The event log does not identify the cause; it shows where and when to investigate.
Time-of-day and shift patterns. Events concentrated in shift-change windows, in the final two hours of a shift, or around break periods suggest fatigue, reduced supervision, or handover failures. These patterns are addressable through scheduling adjustments or pre-shift briefings -- but only if you can see them. A weekly near-miss count by shift reveals this; a monthly summary does not.
Event-class breakdown for root-cause targeting. Forklift-pedestrian proximity events and bypassed-interlock events have different causal structures and require different corrective responses. Treating them as a combined "near-miss count" flattens the signal. A breakdown showing which event classes are rising tells you whether you have a vehicle-management problem, a procedure-compliance failure, or a floor-layout problem -- and which to fix first with limited safety budget.
None of this analysis requires the worker present at an event to file a report. The system observes and logs. The safety team investigates the patterns.
Where vision-based detection does not work
Line-of-sight coverage is a hard constraint. Events behind equipment, in blind corners below camera field-of-view, or in areas without camera coverage produce no record. This is a planning constraint, not a system limitation: camera placement should be designed against your actual blind-spot map before deployment. A coverage audit during the zone configuration stage identifies the gaps.
The system classifies events by observable signals, not by intent. A worker moving quickly through a zone because they are carrying an urgent component produces the same trajectory signature as a worker cutting a corner carelessly. Event records prompt human review; they do not substitute for it.
For outdoor operations across large yards, camera-only detection has coverage gaps. GPS/UWB proximity via the $250 smartband addresses part of the gap for personnel tracking, but multi-hectare outdoor environments fall outside the standard manufacturing-floor deployment pattern. Complex yard-management requirements are better addressed with a combined camera-plus-proximity architecture; contact us for a site-specific configuration.
The system also logs leading-indicator data; it does not prevent the event it detects. A near-miss log that grows in week one of deployment does not mean the floor got less safe -- it means the system is seeing events that were always occurring but never captured. Floor safety improves when the corrective-action cycle acts on the patterns the log reveals. That cycle is a human process.
Cross-industry deployment evidence
Hypernology's 47 production contracts span semiconductor, automotive parts, display panels, PCB, plating, and packaging environments. HyperQ AI Safety's near-miss detection runs the same edge-inference architecture used across those deployments, adapted to safety event classes rather than product defect classes. The trajectory classification model that detects a forklift's convergence path uses the same approach as the model that detects a dimensional deviation in an extruded profile.
The Tier-1 automotive-parts deployment -- 8,000+ product variants, 11,520 units/day across 6 production lines -- represents the closest analog to a high-traffic, multi-vehicle manufacturing floor with frequent pedestrian-vehicle crossing points. The same edge-inference architecture is the basis for HyperQ AI Safety's forklift-proximity event classification.
Cross-industry framing: Hypernology does not have a current named logistics or warehouse safety deployment; the deployment evidence base is manufacturing-floor environments. If your operation is a distribution center or a mixed warehouse-production site, the event classes and zone architecture described here apply directly, but the reference deployment is from a manufacturing context.
For vision-based perimeter monitoring inside robot work envelopes -- where zone design constraints and safety-rated response chains differ from open-floor forklift management -- see 6-axis robot cell safety: vision-based perimeter monitoring inside the work envelope. The HyperQ AI Safety solution page covers the full product configuration and deployment requirements.
Frequently asked questions
Does the system record video continuously, or only when an event fires?
Standard configuration uses event-triggered recording: a rolling buffer is maintained, and a clip is preserved when an event is classified. Continuous full-resolution recording is possible but increases storage requirements substantially. Most facilities use event-triggered recording for near-miss zones and continuous recording only for entry/exit points and high-value equipment areas.
How do we configure event classes for our specific floor layout?
During the 1-hour deployment, the operator defines safety zones on a floor map overlay and assigns event classes to each zone. No custom model training is required for the six standard event classes; custom classes (for example, chemical-spill precursor behavior or specific equipment-state combinations) require additional configuration time and are scoped during the deployment design phase.
What happens when a worker enters an exclusion zone under an authorized permit?
Authorized access is configured per zone during deployment. Workers entering under a permit-to-work are registered as authorized entries; the system distinguishes those from unregistered entries. The configuration integrates with your existing permit-to-work workflow or access control process.
Can we use our existing CCTV cameras for near-miss detection?
Standard CCTV cameras -- rolling shutter, variable frame rate, compressed stream -- typically lack the frame consistency needed for reliable trajectory classification at near-miss timescales. Dedicated cameras are recommended for near-miss detection zones. Existing CCTV infrastructure can serve in supplementary monitoring roles where precision event classification is not required. For the technical reasons behind this distinction, see vision-grade vs surveillance-grade cameras: why a 4K security camera cannot do inspection QC.
Does this system satisfy MOM WSH requirements for near-miss reporting?
There is no current mandatory near-miss reporting requirement under the WSH Act. The system produces structured event records that support your own leading-indicator program and provide an evidence base for investigation when a notifiable dangerous occurrence or injury does occur. Whether the records are useful in a formal MOM inquiry depends on the specific circumstances and your legal counsel's guidance.
What is the minimum camera count for a typical loading bay?
A standard single-bay configuration -- one vehicle entry/exit, defined pedestrian crossing points, clear sightlines -- typically requires two to three cameras for complete zone coverage. High-traffic multi-vehicle bays with multiple pedestrian lanes need four to six cameras. Camera count is confirmed against your floor plan during the deployment design phase, not estimated in advance.
Send us a floor plan covering your forklift aisles, machine exclusion zones, and high-traffic pedestrian crossings. We will map the event classes to your zones, confirm camera count, and outline the expected detection coverage within 5 business days. No contract until the detection specification is agreed and demonstrated on your floor.