An unattended bag is only a problem because of where it is. In an empty yard almost anything stands out; in a rush-hour concourse the same bag is one object among thousands, and that is the case this module was built for.
The Module
Abandoned-object detection is a rule inside ObjectTrack Pro, and the module behind it is a two-stage neural pipeline rather than a motion detector. A first convolutional network annotates objects of known classes on sampled frames and tracks them in between; a second network decides what class each object belongs to; the result is then tested against the rules you drew on that camera. An abandoned bag is one of those rules, alongside people on a rail track and the perimeter work that now has its own page.
Crowd is the hard part, and it is the documented design point. The product guidance is explicit that this module holds up under unfavorable outdoor conditions and in crowds, where a motion-led detector does not: a platform at 08:15 is a moving field of luggage that all looks abandoned for a second or two. The rule therefore works from object persistence inside a region you draw, with a minimum object size and a proximity filter that stops a case being flagged while its owner is standing next to it.
Public space is where it earns its place. The guide names transportation hubs and other public spaces as the setting, and the production estates running it are exactly that: the London Underground, where the platform detects abandoned luggage alongside people falling onto tracks and dangerous platform-edge crowding, and Dublin Airport, where it covers terminals and gate areas at full passenger throughput.
What the trained catalog does not cover, a sentence can. StreamVLM™ adds prompt-defined detectors for objects and conditions no shipped module carries, such as a specific container, a piece of equipment or a local nuisance, with no dataset and no training cycle. StreamVLM is shipping in beta on selected instances.
The Rule
Everything below is drawn or typed per camera, in the module settings, and changed without taking anything offline. There is no site-wide model to retrain when a concourse is rearranged.
The part of the frame to analyze, drawn as a polygon and editable vertex by vertex, plus whether the camera looks at the scene from the top or from the side. It defaults to the whole frame, which is almost never what a concourse wants.
The detector class. Bags, suitcases and comparable items left in the region are what raises the event, rather than any object that stops moving.
Ignore bags closer than suppresses the alert while a person stands beside their case. The exclusion zone is the bag’s own bounding box scaled by a filtration factor you set, so a factor of 2 means a red box twice the size of the yellow one.
Consider abandoned anew after N minutes sets how long one event covers the same object in the same place, so a bag that stays put for an hour does not become an hour of alerts. It is a repeat interval, not a time-to-alarm: IREX publishes no dwell threshold.
A bag or a person has to resolve to at least 1,024 pixels, meaning 32 × 32, before it is a candidate. Vehicles need 2,025, meaning 45 × 45. This is the number that decides how far down a platform one camera reaches.
Stores the motion trajectories of the people and vehicles in the scene so they can be searched in the media player afterwards. It is switched on per camera, and it is what turns a static alert into a line of inquiry.
After the Alert
An abandoned-object alert answers "what" and "where". The investigation needs "who", and the platform gets there in steps rather than in one leap. Saved tracks hold the stored trajectories of everyone who moved through the scene, searchable in the media player, so the approach to the bag can be replayed and the person who set it down identified in the frame. Nothing about that is automatic: the module reports an object, and an operator reads the track back.
A concourse rarely gives you a face. Person re-identification follows an individual from camera to camera on appearance instead, using clothing, accessories, hair or hat color, and ObjectTrack Pro can additionally compute a human silhouette descriptor for every person in the frame, which makes a person searchable by silhouette when nothing else resolves. Both are appearance matching, not identification: they say the same person, not who the person is.
See Person Re-Identification →Once an operator has confirmed the individual, multi-camera object tracking follows them across the camera network on the map and on indoor floor plans, forwards and backwards from the moment of the alert, so the question becomes where they came from as well as where they went. Every step of it runs under the same Case ID and audit rules as any other search: the lawful grounds are recorded before the search runs, and the operator who ran it is on the log.
See Multi-Camera Object Tracking →In the Control Room
Events are filtered onto a named alarm monitor by camera and by module rule, so "Abandoned item" from the concourse cameras becomes one workspace with its own response protocol written into the monitor’s description. Notifications are enabled per monitor and reach responders through the Sover secure messenger with the frame, the camera, the location and a playback link attached.
See Real-Time Alerts & Evidence →The rule name you type becomes the event name, so a post-incident search by event name returns every unattended-item alert on the estate for a period. Combined with saved tracks, that is how an incident review reconstructs a morning rather than scrubbing recordings.
See Ask IREX & Video Search →Any alarm monitor can push its events, with frames, to any system reachable over HTTP, which is how an unattended-item alert lights a public-address workflow, a station control system or an existing incident desk. The REST API covers lists, monitors and users on the same terms.
Requirements
Where It Runs
Unattended items are a public-space problem, so the module lands in the verticals where large numbers of strangers and their luggage move through space nobody controls.
Platforms, concourses and interchanges, including the London Underground deployment where unattended luggage runs alongside track intrusion and platform-edge crowding.
Read More →Terminals and gate areas at full passenger throughput, where an unattended bag alert competes with thousands of attended ones.
Read More →Crossings and inspection halls, where an item left in a queueing area sits alongside watchlist and vehicle work on the same estate.
Read More →Power stations, oil terminals, pipelines and gas facilities run unattended-item detection alongside perimeter intrusion and fire and smoke.
Read More →Entry flow, concourses and stands, where the crowd itself is the reason a left bag is hard to see and important to find.
Read More →Public buildings and federal facilities, where the lobby and the queue outside it are the exposed space.
Read More →Crowded space is the design case, and it is the documented difference between this module and the motion-led one. The rule works from object persistence inside a region you draw rather than from motion across an empty frame, with a minimum object size and a proximity filter tuned per camera. Placement is what decides the result: the requirement is more than 50 pixels per meter in the region that matters.
Not on its own, and no detector should claim to. What the platform provides is the route to the answer: saved tracks store the trajectories of everyone who moved through the scene so the approach can be replayed, person re-identification follows an individual across cameras on appearance when the face is not visible, and multi-camera object tracking then rebuilds their route forwards and backwards in time. An operator confirms the identification, under a Case ID, and the platform asserts no owner by itself.
IREX does not publish a dwell threshold for this module, and we will not invent one. What is documented and configurable is the re-trigger interval, "consider abandoned anew after N minutes", which controls how often the same object in the same place raises a fresh event, and the proximity filter that suppresses the alert while somebody is standing beside the item. The behavior you actually get is tuned on your own cameras during the pilot.
IREX built specialized safety analytics for railways, subways and air transportation, and compiled dedicated training data for these rare incidents by working with transport agencies and generating synthetic 3D environments, which is what made detection reliable in the visual conditions of underground stations. IREX does not publish the size of that dataset, and no per-module accuracy figure is published either: accuracy is benchmarked on your own feeds during the pilot and the measured numbers go into the contract.
That is what StreamVLM™ is for. A plain-language prompt becomes an active detector on the cameras you choose, with no dataset and no training cycle, which covers the long tail of objects and conditions specific to one site. Several prompt-defined detectors can run on the same camera as the trained module. It is shipping in beta on selected instances.
No. ObjectTrack Pro is documented as fixed-camera only, and mobile uploads and drone footage are explicitly out of scope for this module. Uploaded video is analyzed through the Gallery, but with a different set of modules.
It runs on its own and alerts a person. Detections are signals for a human to verify: no response runs autonomously, and the verification and decision are logged against the same Case ID as the detection.
Keep Reading
The concourse at peak is the honest test for this module. Point it at the worst case and judge it there rather than on an empty platform.