Safeguarding passengers on the London Underground. Transit safety is a rare-event problem in hostile visual conditions, so IREX built the training data for those events deliberately and runs the result on one of the busiest underground networks in the world.
In Production
On the London Underground, IREX detects life-threatening situations in real time: people falling onto railway tracks, train surfers, dangerous crowding at platform edges, and abandoned luggage that could indicate a security threat. Reliable detection in the complex visual conditions of underground stations was achieved by building specialized training datasets in collaboration with the transport authority and by generating synthetic 3D environments.
The incidents a network most needs to see are the ones it has almost no footage of. Somebody going onto a track, or riding on the outside of a train, produces very few examples, and collecting more is not something any operator would accept. That is why generic analytics fail here and why the dataset was built rather than gathered: with transport agencies, and in synthetic 3D environments that can stage an event a station will see once a decade.
In Peru, the platform supports public-transport safety across the network, and traffic analytics with two-wheeler detection covers the road side of the same mobility problem.
Bus shelters and transit stops run the same modules on a very different estate, and have their own page.
Scenarios
The incidents a metro or railway operator monitors for, roughly in the order a control room ranks them. Each is a trained detector running on the station cameras, and each links to the page that explains how it works.
Somebody on the running line, walking into a tunnel approach or inside a closed platform end, in the lighting and reflection conditions underground stations actually have. This is the detector a network buys the system for.
See Rail Track Intrusion Detection →Riding on the outside of moving rolling stock. It is one of the incident types with almost no real footage to learn from, which is why the dataset for it was built rather than collected, with transport authorities and in synthetic 3D environments.
See Rail Track Intrusion Detection →Density at the platform edge and in concourses, read while there is still time to hold a flow or open a gate, and a precise count afterwards for the safety review.
See Crowd Management →A bag left on a platform, a concourse or an interchange, raised while the crowd around it is still moving. The rule is set station by station, so a morning peak and an empty midnight platform are not judged the same way.
See Abandoned Object Detection →A firearm drawn on a platform or at a gateline reaches the control room with the frame attached, and the same estate listens for a gunshot, breaking glass or a scream, so the alert does not depend on somebody watching the right screen.
See Weapon & Gunshot Detection →Dwelling where and when dwelling is itself the signal: a closed platform end after the last train, a depot gate, a ticket hall an hour after the station shut. A queue at the morning peak is a different rule, and the same rules cover restricted zones.
See Perimeter Intrusion Detection →Early detection across large volumes, well before point smoke detectors or sprinklers can react. An early-warning layer beside your certified fire system, never in place of it.
See Fire & Smoke Detection →On an estate where nobody is standing at most cameras, a lens that has been sprayed, knocked or turned raises its own alert instead of quietly becoming a blind station.
See Camera Tamper Detection →On the surface network, a vehicle crossing the line when it must not is detected as a railway-crossing violation with an evidence package, on cameras that carry their own installation specification.
See Multi-Camera Traffic Analytics →On Camera
Three frames from three stations. The first is a trained detector; the other two are sentences an operator typed into StreamVLM™, and the band below lists more of them. Two are security events and one is a welfare case, which is the honest shape of transit monitoring: most of what a network needs to see is not a crime. The scenes are illustrative.
A case standing on its own in an interchange at the morning peak, which is the hardest place to see one. The alert reaches the station supervisor with the frame, and the same cameras answer who put it there.
Abandoned Objects →
“Anyone lying on the floor” returns a passenger who has collapsed, someone sleeping rough, and a group who sat down and did not get up. The platform does not decide which is which. It puts the frame in front of the person who can, and the response to one is not the response to another.
Vandalism is expensive because it is found the next morning. A detector written as a sentence fires on the act instead, which is the difference between a repair bill and an intervention. It is shipping in beta on selected instances.
StreamVLM™ →StreamVLM™
Most of what a station calls the control room about is on no vendor’s module list. With StreamVLM™ an operator describes the condition in plain English and the sentence becomes a detector on the cameras the station already has, with its own confidence threshold and alert cooldown, and its alerts routed like any other event. Each row below is a prompt an operator could type this afternoon. StreamVLM™ is shipping in beta on selected instances and runs on a GPU node.
“Detect anyone lying on the floor, day or night.” A passenger who has collapsed on a platform, whether a fall or a cardiac event, and nobody nearby has noticed yet. The alert puts the frame in front of the supervisor who can send someone. It is a welfare check for the operator to triage, not a medical alarm.
“Detect anyone sleeping in the station overnight.” The same posture after the last train is usually a person sleeping rough. The platform reports a person at rest and nothing more; whether station staff or an outreach partner is sent is the operator’s decision, which is why the detector is written as a welfare check rather than a security event.
“Detect graffiti being sprayed in the underpass.” Vandalism is expensive because it is found the next morning. A detector that fires on the act turns a repair bill into an intervention, and nothing in it identifies anyone.
“Alert on any object on the track bed.” A dropped bag, a pushchair, a fallen sign or a bottle on the running line, raised before the next train rather than reported by its driver. The people-on-the-track detector above is trained; this one is a sentence.
“Alert when litter builds up on the platform or a bin is overflowing.” Cleaning dispatched to where it is needed instead of on a fixed round, with a timestamped event for each platform that waited.
“Detect standing water on the platform or in the underpass.” A leak, a burst pipe or a flooded stairwell found before the morning peak reaches it, in time to close the approach.
“Alert when an emergency exit or the foot of an escalator is blocked.” Trolleys, deliveries or a queue that has backed up into the one place it must not, raised while a member of staff can still clear it.
“Detect anyone riding a bicycle or e-scooter on the platform or concourse.” A rule every network has and no camera enforces. The event goes to station staff with the frame attached, and it names no one.
The Modules
The things a network asks its cameras to notice, and the page that explains how each one is detected.
Somebody on the running line, walking into a tunnel approach, or riding on the outside of a train. This is the detector a metro network buys the system for.
Read More →Density read across a platform or a concourse while there is still time to hold a flow or open a gate, and a precise count afterwards for the safety review.
Read More →The hard version of unattended luggage: one object standing still in a crowd that is not. The alert carries the frame, and the archive answers who left it there.
Read More →A firearm drawn in a ticket hall or on a platform, and the gunshot, breaking glass or scream the same estate listens for, reaching the control room with the frame attached rather than as a call from the platform.
Read More →A no-loitering zone with a time limit on a closed platform end, a depot gate or a ticket hall after hours, and the restricted zones the same rules cover.
Read More →Large volumes where a point detector in the ceiling reacts late or not at all. An early-warning layer beside your certified fire system, never in place of it.
Read More →A passenger on the floor, an overnight stay, graffiti, debris on the track, a bin overflowing: the scenarios above, described in a sentence instead of waiting for a vendor module. Shipping in beta on selected instances.
Read More →A sprayed lens, a knocked mounting or a slow slide into unusable video, raised as an event. On an unstaffed platform that is the difference between a fault and a blind station.
Read More →Railway-crossing violations on the surface network, detected with an evidence package on cameras that carry their own installation specification.
Read More →By building the data rather than waiting for it. IREX compiled specialized training datasets for these rare incidents in collaboration with transport agencies and by generating synthetic 3D environments, which is what made reliable detection in complex underground conditions achievable. We publish no dataset size and no accuracy figure: what we commit to is measuring detection on your own cameras during the pilot, against criteria agreed in writing.
That is the condition it was built for. The London Underground deployment is the reference, and the dataset work with the transport authority and with synthetic 3D environments was aimed specifically at underground visual conditions.
No. An edge server runs the analytics locally and sends only events rather than raw video, keeps recording offline when the link drops, and synchronizes automatically on recovery. High-availability edge is available where a site cannot afford to go dark.
Yes, through crowd analytics: real-time dangerous-crowding alerts plus precise post-event counting, estimating up to 3,000 people in a defined region of a camera view.
No, and it does not try to. The prompt-defined detector reports a person on the floor and puts the frame in front of the operator, who decides whether to send station staff, an outreach partner or an ambulance. That is deliberate: IREX products are expressly not for life-saving or emergency systems, and the decision stays with a person. StreamVLM™ is shipping in beta on selected instances.
People on the track, train surfing, dangerous crowding, abandoned items, weapons, loitering, fire and smoke and camera integrity are trained modules in production. The scenarios written as sentences, from a passenger on the floor to a bin overflowing, are StreamVLM™ prompt-defined detectors: they ship in beta on selected instances, run on a GPU node, and whether a specific deployment is in scope is confirmed with IREX engineering rather than promised in advance.
No. Reusing the existing fleet is the standard deployment model: any camera that supports ONVIF or streams RTSP with H.264 or H.265 on a static IP can be connected, and each module then sets its own resolution and placement requirements, which a site survey confirms. Note that connection is not plug-and-play: a qualified network engineer configures each camera and router.
You do. Customers retain 100% ownership and control of their data, and IREX neither owns nor accesses customer video, events, logs, watchlists, or floor plans. On-premises deployment is the data-residency mechanism.
With a pilot: one site, three to five use cases, six to twelve weeks, and two or three measurable success criteria agreed in writing before it begins. Accuracy is benchmarked on your own cameras, and the measured numbers go into the contract.
Keep Reading
A single station, instrumented properly, tells you more about a network program than a paper study of all of them.