Generate a loooooooooooot of data with a little set of files and train a model with it

cd model
python keras_model.py
python sklearn_model.py

• in keras_model.py the genData() function allows to generate tons of data from data/train and data/validation folders, you can put what ever you want in it, make sure to rename properly the folder name. In this example, it trains digital number (unlike the MNIST dataset).
• run keras_model.py will generate lot of data and train the model from 0 to 9. and then you can detect digits from the traffic light if your country has it.

Traffic light detection with Yolov3, then process direction, color and digits with trained model

some comparison between keras and sklearn model profile:

One could find running the keras model is too much memory consuming, there are other methods to tune the tf backend, here is an example using scikit-learn.

hyperParametersTuning.png

deeplab segmentation mask, lane boundaries extraction, ransac to eliminate outliers, cv draw line