Traffic Sign Recognition

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

Data Set Summary & Exploration

1. Dataset Summary

I used the pandas librar=y to calculate summary statistics of the traffic signs data set:

• The size of training set is 34799
• The size of test set is 12630
• The size of the validation set is 4410
• The shape of a traffic sign image is 32x32x3
• The number of unique classes/labels in the data set is 43

2. Classes/Labels Distribution

In the bar charts bellow the X-axis show the code for the class (the meaning of the codes can be found here) and the Y-axis is the amount of images for that class. It is paramount for a good training/testing/validation that the distribution of the image classes follows the same pattern.

Design and Test a Model Architecture

As a first step, I decided to convert the images to grayscale because first it decreases the amount of data being managed by a a factor of 2 (the images goes from 32x32x3 to 32x32x1). Second, I found out that using gray images leads to a better model accuracy, what was also seem in this article by Pierre Sermanet and Yann LeCun. I suppose it is somewhat intuitive that colors are not important for this aplication, by eye we can identify the signs with color as well as in grayscale.

Here is an example of a traffic sign after grayscaling.

Normalizing the data is a good practice because it keeps numerical stability of the algorithm and having features with a 0 mean and equal variance gives a Well Conditioned Problem for the optimizer. The normalization formula is diplayed bellow.

$$normalized_value = pixel_value - 128 / 128$$

By generating adittional data I was able to boost the quality of the model. Adittional 5 images were generated for each image in the training set by aplying a random perturbation to it.

Here is an example of an original image and an augmented image:

The difference between the original data set and the augmented data set

• Images were translated by a random amount of pixels vertically and horizontaly in the interval [-2, 2]
• Images were tilted by a random degree amount in the interval [-15, 15]

####2. Describe what your final model architecture looks like including model type, layers, layer sizes, connectivity, etc.) Consider including a diagram and/or table describing the final model.

My final model (LeNet) consisted of the following layers:

Layer	Description
Input	32x32x1 gray image
Convolution 5x5	1x1 stride, valid padding, outputs 28x28x6
RELU
Max pooling	2x2 stride, outputs 14x14x6
Convolution 5x5	1x1 stride, valid padding, outputs 10x10x16
RELU
Max pooling	2x2 stride, outputs 5x5x16
Fully connected	input 400, output = 120
RELU
Dropout	keep probability = 0.5
Fully connected	input 120, output = 84
RELU
Dropout	keep probability = 0.5
Fully connected	input 84, output = 43
RELU
Softmax

To train the model, I used the Adam Optimizer and used its default learning rate (0.001). Using a NVIDIA Titan X to train the model allowed me to use a big batch size, however in my Jupyter notebook the batch is set for 128 (someone might try on an older GPU). No significant increase in the performance were observed above 40 epochs.

My final model results were:

• test set accuracy of 95.3%
• validation set accuracy of 96.7%

Testing random images found on Google

Speed limit (70km/h)

Probability	Prediction
1.0	Speed limit (70km/h)
5.1180215e-35	Speed limit (30km/h)
Almost 0.0	Speed limit (20km/h)
Almost 0.0	Speed limit (50km/h)
Almost 0.0	Speed limit (60km/h)

Children crossing

Probability	Prediction
0.73999822	Children crossing
0.25391123	Bicycles crossing
0.0060873665	Road work
2.5763277e-06	Beware of ice/snow
5.7231995e-07	Slippery road

Road work

Probability	Prediction
1.0	Road work
Almost 0.0	Speed limit (20km/h)
Almost 0.0	Speed limit (30km/h)
Almost 0.0	Speed limit (50km/h)
Almost 0.0	Speed limit (60km/h)

Stop

Probability	Prediction
1.0	Stop
2.1838753e-08	No entry
6.8094623e-09	Keep right
1.3789767e-09	Turn right ahead
1.8331527e-11	Yield

Double curve

Probability	Prediction
0.9999603	Double curve
3.9675648e-05	Right-of-way at the next intersection
3.0263076e-09	Wild animals crossing
2.3719895e-09	Road work
2.2993707e-09	Slippery road

Go straight or right

Probability	Prediction
0.99594325	Go straight or right
0.0034704579	Keep right
0.00051441899	Turn left ahead
6.8129419e-05	Ahead only
2.9697715e-06	Yield

Roundabout mandatory

Probability	Prediction
0.99163145	Roundabout mandatory
0.0081623141	Speed limit (100km/h)
0.00019325051	Priority road
9.4156412e-06	No entry
2.9617011e-06	Keep left

The model seems to work very well for the images selected other than the "Children Crossing" image, which accordingly to the prediction has a 25% chance of being "Bicycles crossing". The could be a raction of many factors as the lack of smaller amount of "Children Crossing" provided for training the model when comparing to the other classes as well to the fact that these two classes seems to have a lot of pixels in common.