Safe Rail
Frame by frame track bed and ballast segmented from a forward-facing camera.
- Client
- Confidential
- Year
- 2025
- Category
- Computer Vision
- Role
- Design + Build
- Timeline
- 4–6 weeks
The problem
Track inspection is a person walking the line or riding a slow trolley. It’s expensive, infrequent, and only as consistent as the inspector’s attention that day.
What we built
A semantic-segmentation model runs on a forward-facing camera feed, labelling rails, sleepers and ballast in every frame and flagging where the ballast profile or the bed looks wrong. Trained on annotated footage from the corridor it monitors, so it holds up in real lighting and real motion blur.
The result
Track condition is captured continuously from a camera already on the vehicle, instead of once a quarter on foot. Flagged sections go to a human for review with the frame and timestamp attached.
- rail, sleeper and ballast segmentation
- every framerail, sleeper and ballast segmentation
- detection confidence on the monitored corridor
- 0.9+detection confidence on the monitored corridor
- runs on a camera already mounted for the run
- on-vehicleruns on a camera already mounted for the run
Built with
- Python
- YOLO
- Semantic segmentation
- OpenCV
- Data annotation
We deliver what we commit.
Tell us what you're trying to build.
We'll come back within 24 hours with honest feedback on scope, timeline and cost, whether or not we turn out to be the right fit.