Traffic Sign Recognition Project

The Project

The goals / steps of this project are the following:

• Load the 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

Data Set Summary & Exploration

1. Dataset summary

I used the python and pandas library to calculate summary statistics of the traffic signs data set:

Description	Value
Input	32x32x3 RGB image
The size of training set is	34799
The size of the 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. Dataset distribution

Here is an exploratory visualization of the data set. It is a bar chart showing how the data is distributed over the classes

Design and Test a Model Architecture

1. Preprocess

As a first step, I decided to convert the images to grayscale to reduce the amount of inputs by reducing channels of the image.

Here is an example of a traffic sign image before and after grayscaling.

I normalized the image data to maintain data near to 0 and simplify the training process

The difference between the original data set and the augmented data set is the following ...

2. Architecture

My final model consisted of the following layers:

Layer	Description
Input	32x32x1 RGB image
Convolution 5x5	1x1 stride, valid padding, outputs 28x28x16
RELU	droupout prob 0.8 after if training
Max pooling	5x5 stride, outputs 14x14x16
Convolution 5x5	1x1 stride, valid padding, outputs 10x10x64
RELU	droupout prob 0.8 after if training
Max pooling	5x5 stride, outputs 5x5x
Convolution 2x2	1x1 stride, valid padding, outputs 4x4x120
RELU	droupout prob 0.8 after if training
Max pooling	2x2 stride, outputs 2x2x120
Flattening	output 480
Fully connected	output 120
Fully connected	output 43 (classes)
Softmax

3. How you trained my model.

To train the model, I used an adam optimizer and apply 128 samples batch size. To improve the accuracy I repeat the training fo 30 ephocs with a 0.001 lerning rate value.

4. Approach

My final model results were:

• training set accuracy of 0.996
• validation set accuracy of 0.957
• test set accuracy of 0.946

The first architecture was the LeNet but was not enough to get more than 0.89 accuracy. I tried to use the YUV color space as in this paper. But with my actual architecture was still under 0.85~0.87 of accuracy. By adding another convolutional layer the accuracy increase. After that I added a dropout to make the recognition more flexible. The accuracy rise the 0.93 accuracy araound the 20th ephoc but i decided leave 30 to consolidate the pattern.

Test a Model on New Images

1. Choose five German traffic signs on the web

Here are five German traffic signs that I found on Google Maps moving around Hamburg by taking sreenshots.

The prediction was pretty accurate.

2. Model's predictions

Here are the results of the prediction:

Image	Prediction
Yield	Yield
Keep right	Keep right
Road work	Road work
No entry	No entry
Speed limit (30km/h)	Speed limit (30km/h)

The model was able to correctly guess 5 of the 5 traffic signs, which gives an accuracy of 100%.

3. Probabilities of predictions

The code for making predictions on my final model is located in the 11th cell of the Ipython notebook.

For the first image, the model is has no doubt, and also others images returns similar values on probabilities.

Probability	Prediction
1.000000	Yield
0.000000	No vehicles
0.000000	No passing for vehicles over 3.5 metric tons
0.000000	Right-of-way at the next intersection
0.000000	Ahead only

But let's se the last one that introduce a little uncertainty on the value of the speed limit.

Probability	Prediction
0.911139	Speed limit (30km/h)
0.077434	Speed limit (50km/h)
0.005195	Speed limit (20km/h)
0.001936	Vehicles over 3.5 metric tons prohibited
0.001237	Speed limit (80km/h)