IREX
EN

StreamVLM

StreamVLM™ creates a working video detector from a sentence. A condition described in plain text becomes an active detector on selected cameras, with no dataset collection, no labeling, and no model training.

Availability. StreamVLM is shipping in beta on selected instances. The capability is included in the standard IREX price; the only additional cost is a local GPU node for Vision-Language Model inference where a deployment requires one. To confirm whether a specific deployment is in scope, contact the IREX representative for the region.

The problem it solves

A conventional analytics module is a trained model. Adding one means collecting footage of the condition, labeling it, training, validating, and shipping a release. That effort is justified for weapons, faces, plates, and fire, where the condition recurs across every deployment and the stakes are high.

It is not justified for the long tail: standing water in one underpass, illegal dumping behind one building, uncleared snow on one fire exit, graffiti on one wall. These conditions are specific to a municipality, and there are hundreds of them. StreamVLM covers that tail, because describing a condition costs a sentence instead of a development cycle.

How it works

How a StreamVLM prompt becomes an alert FROM PROMPT TO ALERT 1. Prompt Written once, in plain text: "a person lying on the ground" 2. Frames Sampled from the live stream of the selected camera channel. 3. VLM A Vision-Language Model reads image and text together, then scores each frame. 4. Threshold Per-prompt confidence and alert cooldown decide what qualifies. 5. Event Real-time alert, searchable record, dashboard data. The prompt, the evaluation, and every resulting event are written to the audit log Inference runs on a GPU node inside the customer network. Frames and results do not leave it.
A prompt is configured once and then evaluated continuously against sampled frames. Everything the detector produces enters the same event pipeline as any other analytics module.

What a detector looks like

A StreamVLM detector is a named set of prompts with its own settings, applied to selected camera channels. Each prompt is independent:

Setting What it controls
Prompt text The condition to detect, described in plain language.
Confidence threshold How certain the model must be before an event is raised. A lower threshold catches more and reports more false positives; a higher one does the reverse.
Alert cooldown The minimum interval between repeat alerts for the same condition on the same channel, so a persistent condition does not flood the operator.
Event type How the resulting event is classified for search, alarm monitors, and reporting.
Cameras Which channels the detector runs on.

A single camera channel supports up to 100 prompt-defined detectors at once. A station camera can therefore watch simultaneously for a person on the tracks, platform overcrowding, an unattended bag, smoke, fresh graffiti, and flooding at platform level. Prompts are added and adjusted without taking the system offline.

Example prompts

Scenario Prompt
Medical emergency in a public space a person lying on the ground
Flooding in an underpass standing water or flooding in the underpass
Infrastructure damage damaged or broken fencing along the perimeter
Illegal dumping debris or trash dumped outside a container
Vandalism fresh graffiti on a wall
Winter hazard uncleared snow on a sidewalk or fire exit
Restricted area a vehicle parked on the sidewalk
Signage and markings a container number or warning sign visible on the vehicle

Where StreamVLM fits alongside the other modules

StreamVLM complements the trained modules rather than competing with them. The two work at different frame rates and answer different questions.

Trained modules StreamVLM
Best at Tracking and identifying people, vehicles, and objects continuously Recognizing a described scene or condition
Frame rate High, every frame, so an object can be followed Sampled frames, sufficient for a condition that persists
New capability requires A dataset, training, and a release A sentence
Typical use Watchlist matches, plate reads, intrusion, counting, traffic enforcement Municipal conditions, site-specific hazards, one-off monitoring needs

StreamVLM events enter the same pipeline as any other event. They work with alarm monitors, appear in event search, reach the field through Sover, and export to Apache Superset dashboards.

Accountability

A detector that anyone can define in a sentence needs the same oversight as one that took a year to train, and it is governed by the same mechanisms:

  • Prompts are recorded. Creating, editing, and activating a detector are logged actions, attributed to the operator who performed them. What the system was told to look for is therefore always reconstructable.
  • Events are auditable. Every alert carries its channel, timestamp, confidence, and the prompt that produced it, which is what makes a detector's behavior reviewable after the fact.
  • Searches still require a Case ID. Retrieving StreamVLM events from the archive is an event search, so the rules in Case ID and accountability apply unchanged.
  • Permissions still apply. Who may create prompts, on which cameras, and who may see the resulting events is governed by role-based access control.
  • Prompts do not widen the platform's scope. A prompt describes a scene or a condition. It does not create a new identification capability, and it cannot be used to recognize or track individuals who are not pre-registered. See Engineered for ethics.

Related topics