Industrial AI
Computer vision for in-plant traffic safety
Heavy vehicles share plant roads with people. A speed limit only works if someone can see when it is broken. A camera and a trained model now do that every hour of the day.
Problem
Speed limits on internal roads were enforced by occasional checks. There was no record of how often they were exceeded, by which kind of vehicle, or at what time of day.
Approach
Use an existing site camera that overlooks a straight road section. Train a YOLO detection model on footage from that camera, so it recognises the vehicle type that matters most rather than every moving object.
Mark two measurement lines a known distance apart on the road in the camera view. Speed is the distance divided by the time a tracked vehicle takes between them.
Save a snapshot and a record for each event, show live video with the measurement lines and speed on a web screen, and send a summary by email.
Run the service unattended: a watchdog task starts it at boot and restarts it if it stops, with a maintenance switch to pause it on purpose.
Architecture
One camera, one GPU workstation, one review screen.
Result
The road is now monitored continuously, and every event has a time, an estimated speed and a picture. Safety discussions start from records instead of impressions.
The main lesson: in a plant, most of the effort goes into network access to the camera, sharing one GPU with other services, and making the service recover by itself after a restart that nobody planned.
Stack
Python, YOLO detection with a custom-trained model, object tracking, a small web server for the review screen, scheduled tasks for unattended operation.