This is a console application that aim to detect the traffic signs encountered in camera.
This project is built using keras and openCV libraries.
This is a Convolutional neural network based approach over traditional model. This system approaches the problem by employing 3 layered neural network and passing the preprocessed images in them through pipeline.
The model should detect and process the traffic signs of 43 different classes.
It should return the class and name of the input (traffic sign) provided.
Images are 32 (width) x 32 (height) x 3 (RGB color channels).
Training set is composed of 22271 images.
Validation set is composed of 5568 images.
Test set is composed of 6960 images.
There are 43 classes (e.g. Speed Limit 20km/h, No entry, Bumpy road, etc.).
Exploratory Visualization of Images
A difference can be observed in the no. of different classes of images, this change in number also results in different accuracy of each class.
Network has 3 Convolutional Layers – kernel size: 3 X 3 Activation function – ReLU (followed by 2 X 2 maxPooling operation) Last 3 layers are connected producing 43 classes (using SoftMax activation fn.)
It is observed that after 8 epochs the loss and accuracy becomes constant and the result is 99% accurate.