Underground Changes Every Safety Rule You Know

Singapore is building the longest fully underground railway in its history. The Cross Island Line will run more than 50 kilometres across 21 stations, and where it passes beneath the Central Catchment Nature Reserve it sits roughly 70 metres down. LTA describes that depth as the height of a 25-storey HDB flat, inverted, below the ground you are standing on.

Most safety thinking is built for sites you can see from the road. Underground construction is not that. It changes the shape of nearly every hazard on a standard risk register, and it changes what a safety team can realistically observe.

We work with contractors delivering underground rail packages. This is what we have learned about how those hazards behave once the work goes below ground, and what a monitoring layer has to do differently to be useful there.

Work at height stops meaning falling off things

On a building site, work at height is mostly about falling from a structure. Underground it inverts. The exposure is falling into a shaft.

Access shafts on a deep line can run tens of metres. Around them sit temporary decks, void openings between levels, launch chambers and formwork edges that change position weekly as the works advance. The fall is not from a visible edge at the top of something. It is into an opening in the floor you are walking across.

That matters for detection because the visual problem is different. There is no sky, no horizon, no consistent daylight. Cameras work in artificial light, through dust and water vapour, against dark backgrounds, often looking down rather than across. A model tuned on sunlit rooftop footage does not transfer. Detecting a worker near an unprotected shaft edge at 40 metres depth is a harder computer vision problem than detecting one on a scaffold, and it needs to be treated as one.

Proximity to machinery matters more because there is nowhere to go

On an open site, a worker who notices a reversing excavator can step aside. In a tunnel there is no aside.

A tunnel bore is a corridor. Segment erectors, muck trains, locomotives, grout plant and the boring machine itself operate in a space with no lateral escape route. The clearance between plant and the tunnel wall is often measured in centimetres. When a worker and a machine occupy the same section of that corridor, the options for avoidance are close to zero.

This is why plant proximity detection earns its place underground more clearly than anywhere else. On the surface it reduces a risk. Underground it substitutes for a margin that physically does not exist. The alert has to reach someone who can stop the movement, and it has to arrive in seconds, because the geometry gives no second chance.

Exclusion zones move, so static signage stops working

Underground rail work runs a rotating set of temporarily lethal areas. A cutterhead intervention. Segment handling. Grouting operations. A shaft under a live lift. Each of these makes a defined area unsafe for a defined window, then releases it.

Physical barriers and printed signage are built for permanence. They are slow to move, and once a zone has been safe for a week people stop reading the sign. Zone rules that can be redrawn per camera, per shift, per activity match how the work actually behaves. A restricted area that exists only while the activity is live is enforced accurately, and stops being background noise the rest of the time.

Missing barricades are a state change, and state changes are what cameras catch

Almost nobody removes edge protection intending to leave it off. It comes off for access, for a delivery, to swing a load through, and the intention is to put it back.

The gap between removal and replacement is where people fall. It is also nearly invisible to periodic inspection, because a walkthrough sees a snapshot. A barricade missing for ninety minutes between two inspections leaves no trace in any record.

Continuous monitoring is well suited to this specific failure because it detects a change of state rather than a condition. The question is not whether a barricade exists on the drawings. It is whether the opening is protected right now.

Lone workers underground lose the tools built to protect them

Long tunnel drives, cross-passages, plant rooms and night possessions all produce work where somebody is alone for a period. That is normal and often unavoidable.

The problem is that most lone worker technology assumes connectivity. Phone-based check-in apps, cellular man-down alarms and GPS tracking degrade underground precisely where the exposure is highest. The tools designed for the risk stop working at the point the risk becomes real.

A camera-based layer running on site infrastructure does not carry that dependency. It also does not require the worker to carry, charge or activate anything, which removes the most common reason lone worker systems fail in practice, which is that people do not use them.

The other things worth watching underground

The five above are the ones people ask about. In practice an underground programme runs a wider set:

• Confined space entry, including whether entry matches the permit and whether the attendant is actually present
• Hot work and fire watch, where the watch period after work stops is as important as the work itself
• Lifting operations, particularly loads passing over shafts and anyone standing under a suspended load
• PPE by zone, since requirements differ between the surface compound, the shaft and the tunnel face
• Housekeeping, because slips and trips underground happen on wet, cluttered, poorly lit walkways
• Vehicle and pedestrian separation at the surface interface, where site traffic meets public roads
• Excavation and shoring, including edge proximity around open cuts
• Smoke and fire detection, where early warning matters disproportionately given restricted egress
• After-hours access, since an unattended shaft is both a safety and a security exposure

Not every site needs all of these. The point is that the hazard set underground is broader than the surface equivalent, not narrower, and coverage should reflect that.

What a monitoring layer actually gives you on a programme like this

The value is not that a camera spots something a person could not. A competent supervisor spots all of it. The value is in the arithmetic of coverage.

Continuous instead of sampled. A safety officer running six walkthroughs a day covers a fraction of a large site at any moment. Cameras that are already installed watch every connected area, every shift, including the night shift when supervision is thinnest.

Evidence attached to every observation. Each detection carries the video. That removes the argument about what happened and turns a toolbox talk from an assertion into a replay. It also gives an incident investigation something better than recollection.

Intervention while it still matters. An alert in seconds to the person who can act is a different thing from a report at the end of the week. Most of the value sits in the gap between those two.

Patterns across sites, not just within them. On a multi-package programme the same failure often recurs across several worksites with different contractors. Nobody sees that from inside one site. Aggregated observations make a repeating systemic issue visible, which is the difference between fixing an instance and fixing a cause.

No new hardware in most cases. These sites already have CCTV for security and progress monitoring. Running detection on existing RTSP and ONVIF cameras avoids a second install programme in an environment where every penetration, cable route and power supply is already contested.

Records that suit the reporting you already do. Every observation is logged and searchable, which supports internal review and external reporting without a separate data-gathering exercise.

What we would tell another team going underground

Three things.

Do not assume surface-tuned detection transfers. Budget for the models to see tunnel conditions before you judge accuracy.

Decide early who receives an alert and what they are expected to do. A detection that reaches nobody with authority to stop work is an observation, not a control.

Expect the alert volume to be wrong at first, in both directions. Tuning the balance between noise and coverage is the work, and it takes real footage from the real environment to get right.