Platform Ends
Select platform-end and track-approach cameras with an unobstructed view of the dangerous zone.
Person on Railway Tracks Detection for Railway Review
Someone on the track area needs attention before the operator decides what is happening. StreamVLM™ can surface the visible condition from selected station cameras.
The Visible Condition
Give the Operator a Frame to Review
The condition in scope is a person standing, walking, or climbing down within the railway track area. It is a visible scene detail, not a diagnosis, a conclusion about intent, or an automatic decision.
In StreamVLM™, an authorized operator writes a prompt such as “Alert on anyone on the railway track.” The model evaluates selected live frames against that prompt and sends a match, camera, location, and time into the IREX event workflow for human review.
StreamVLM™ runs on a GPU node in the customer instance. Each prompt has its own confidence threshold, alert cooldown, event type, and camera selection.
This prompt-defined detector can be evaluated alongside the Track Intrusion Detection trained-module workflow. The two approaches have different frame-rate and operational-fit considerations, so a pilot should establish which is appropriate for the site.
Explore StreamVLM Prompt DetectorsWhere It Helps
Set the Prompt Where the Condition Matters
Track Approaches
Let control-room staff confirm the person’s position and nearby train movements.
Open-Cut Stations
Use the station’s existing escalation path, rather than inferring circumstances from the alert.
From Prompt to Action
A Review Workflow for This Condition
- 01
Define the Track Area
Write a concise prompt for a person standing, walking, or climbing down within the railway track area, then select the cameras and times where the condition warrants attention.
- 02
Check the Live Platform and Track Context
Open the event with its camera, location, time, and available video context. Confirm what is visible before treating the match as actionable.
- 03
Notify the Appropriate Rail-Control or Station Team
The authorized team applies the local procedure. StreamVLM supplies a review cue; it does not decide the response.
Limits to Test
Validate It on the Actual Cameras
- Scene Context
- A person on the platform edge is different from a person in the track area.
- Camera Fit
- The tracks, platform boundary, and person need to be visible in the camera view.
- Human Decision
- The frame cannot establish motive, distress, or the safest response.
- Timing
- Track intrusion can be time-critical. StreamVLM evaluates sampled frames, so the rail operator must validate the camera view and response timing during a site-specific pilot.
FAQ
Can the prompt tell why a person is on the track?
No. It only surfaces the visible position. Staff must review the live situation and use their established railway safety procedure.
Can StreamVLM confirm person on railway tracks?
No. It reports a match to the visible condition in the prompt. An operator reviews the frame and local context before deciding what it means.
How should a team tune the prompt?
Start with representative cameras, define what makes a useful alert, then adjust the camera selection, confidence threshold, and alert cooldown during a camera-specific pilot.
Is there a published accuracy figure for this use case?
IREX does not publish a universal accuracy figure for this prompt. Image quality, camera angle, lighting, occlusion, and the local definition of a useful alert all affect results.
Pilot a Track-Area Review Workflow
Start with one platform-end view and agree who receives the alert, who confirms it, and how the station response is coordinated.