Stolen carsfrom above

Insurers pay out when a car is stolen and rarely recover the vehicle. The longer it takes, the lower the odds, and today the search depends on ground patrols, which cover very little area per hour. The question behind this project was whether a drone sweeping car parks and lots could do that search usefully.

The system

It is several vision models working in a chain, not one. The first locates vehicles in the drone frame. The second isolates the plate region inside each vehicle found. The third reads the characters. Each stage’s output is the next one’s input, and each can fail in a different way.

The readings are then matched against a stolen-vehicle database, and anything that matches is flagged.

Why split it into stages

A single model trying to find and read plates directly in the full drone image has to deal with plates occupying few pixels in the full frame. Splitting it into vehicle detection → plate crop → reading means each model receives an image already framed on the problem it solves, and the crop reaches the OCR enlarged.

What stayed out of reach

Flight altitude and camera angle change the difficulty of the problem, and that was the hardest variable to control in testing.