Traffic Sign Recognition

Writeup

Build a Traffic Sign Recognition Project

The goals / steps of this project are the following:

Load the data set (see below for links to the project data set)
Explore, summarize and visualize the data set
Design, train and test a model architecture
Use the model to make predictions on new images
Analyze the softmax probabilities of the new images
Summarize the results with a written report

Rubric Points

Writeup / README

Data Set Summary & Exploration

1. Basic summary of the data set

The code for this step is contained in the 3rd code cell of the IPython notebook.

I used python and numpy libraries to calculate summary statistics of the traffic signs data set:

The size of training set is 34799
The size of validation set is 4410
The size of test set is 12630
The shape of a traffic sign image is (32, 32, 3)
The number of unique classes/labels in the data set is 43

2. Visualization of the dataset.

The code for this step is contained in the 5th and 6th code cell of the IPython notebook.

Here is an exploratory visualization of the data set and a bar chart showing the distribution of sign types

Design and Test of the Model Architecture

1. Description of Preprocessing of the Image Data.

The code for this step is contained in the 20th code cell of the IPython notebook.

As a first step, I decided to convert the images to greyscale because this would speed up training time without decreasing prediction quality.

Here is an example of a traffic sign image after greyscaling.

As a last step, I normalized the image. So that all values are between 0. and 1.

2. Augmentation, Training, Validation and Testing Data.

I used the training, validation and test data as it was provided.

The size of training set is 34799
The size of validation set is 4410
The size of test set is 12630

to increase the number of training data and to balance out the vastly uneven distribution of sign types in the training data, I generated additional data from the existent images.

I generated new data (especially for the signs with low occurance) by

rotation by -1, -2, 1, and 2 degrees around the center of the images
zoom in by 5% and 10%

The data augmentation is realized in the 11th code cell of the IPython notebook.

This step increased the size of the training set from 34799 to 302864 (â‰lmost 10x!) images and balanced out the distribution of different sign type quite good

The images in the validation set and test set remained unchanged.

3. Model Architecture

The code for my final model is located in the 24th cell of the ipython notebook.

My final model consisted of the following layers:

Layer	Description
Input	32x32x1 greyscale image
Conv1 3x3	1x1 stride, same padding, outputs 32x32x32
RELU
Max pooling	2x2 stride, outputs 16x16x32
Conv2 3x3	1x1 stride, same padding, outputs 16x16x64
RELU
Dropout	0.7
Max pooling	2x2 stride, outputs 8x8x64
Conv3 3x3	1x1 stride, same padding, outputs 8x8x128
RELU
Dropout	0.7
Max pooling	2x2 stride, outputs 4x4x128
Fully connected FC1	1024
RELU
Dropout	0.7
Fully connected FC2	512
RELU
Dropout	0.7
Fully connected FC3	43
Softmax

4. Hyperparameters

The code for training the model is located in the 27th cell of the ipython notebook.

To train the model, I used:

Adam Optimizer
Learning Rate of 0.0001
batch size of 128
50 Epochs

5. Approach for Finding a Solution.

The code for calculating the accuracy of the model is located in the 27th and 28th cell of the Ipython notebook.

I first tried the architecture like it is given in the lecture notes. Validation accuracy maxed at 0.91 and test accuracy was at 0.83.

To increase the over all accuracy I introduced an additional convolutional layer and an additional fully connected layer. The gap between validation accuracy and test accuracy indicates overfitting. To reduce this gap I introduced dropouts (0.7) after conv2, con3, fc1 and fc2

The learning rate of 0.001 and 0.00001 didn't work - with these rates the validation accuracy remained at 0.05 through all epochs.

I generated as much new training data as possible that threw no memory errors when I started the training.

The most important measures were to introduce dropout and to augment the training data to increase training samples and to balance the occurance of sign types.

My final model results were:

validation set accuracy: 0.958
test set accuracy: 0.941
accuracy on my traffic signs: 1.0

The test accuray (0.941) is very near the validation accuracy (0.958). That proves that the model works well, generalizes well and does not overfit.

Testing the Model on New Images

1. New Images

Here are five German traffic signs that I took in Munich: