Putting computer vision on a nation's roads, before it was fashionable
What it took to run phone and seat-belt detection at national scale in 2018: hardware, data and the review loop that made it trustworthy.
· 6 min read · Cozyplus Technologies
In 2018 there were no off-the-shelf models for spotting a phone in a driver's hand through a windscreen at night. We trained our own, on regional footage, with glare, tinted glass and a thousand vehicle shapes.
The part nobody talks about
The model was the smaller problem. The larger one was building a system that a human reviewer could trust: every detection carried its frame, the model version and a confidence score, and a person confirmed each case before it became a fine. That loop is why the system is still running.
What we would do the same today
Deploy on the client's own GPU servers with no external connectivity. Re-evaluate the models on fresh samples every month. Treat the audit trail as the product, not a feature.