plateExtractor

Extracts vehicle plates from a video stream via CNN using Keras.

Usage

pip3 install -r requirements.txt
./main.py {pathOrUrlToVideo}

The model

You can generate any new model on your own by using train.py under scripts folder. Current model uses:

• convolution(32, kernel=5)
• convolution(64, kernel=3) with maxpooling, dropouts and batch normalization.

Note that this is trained and tested on Turkish vehicle plates.

Scaling

Current state seems to be optimal for 1920x1080 frames and it may not work well for other resolutions yet, though it should be easily doable.

TODO

You are invited to create and fix issues.

• Make sure it works for all frame resolutions (plateExtractor:sobel_fill_dims and various hardcoded values such as plateExtractor:minContourArea)
• After reading plates, convert ambigious numbers to letters or vice-versa to fit the plate regex pattern (8 -> B, 0 -> D, I -> 1 ...)
• Refactor