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

---

Writeup

Link to project code

Data Set Summary & Exploration

1. Basic summary of the data set.

I used the numpy library to calculate the summary statistics of the traffic signs data set:

• The size of the training set is 34799
• The size of the validation set is 4410
• The size of the 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. Exploratory visualization of the dataset.

Here are the examples of the images from the training set:

Here is an example of classes distribution from the training set:

Design and Test a Model Architecture

1. Data preprocessing

As a first step, I decided to convert the images to grayscale. It reduces the dimension of input and makes normalization of the image easier.

At the next step, I am normalizing the data in order to decrease complexity and potentially increase network accuracy and learning speed.

Raw	Gray	Normalized

2. Model structure

My final model consisted of the following layers:

Layer	Description
Input	32x32x1 RGB image
1. Convolution 3x3	1x1 stride, valid padding, outputs 28x28x6
RELU
Max pooling	2x2 stride, outputs 14x14x6
2. Convolution 3x3	1x1 stride, valid padding, outputs 10x10x16
3. Fully connected	input 5x5x16, output 400
RELU
4. Fully connected	imput 120, output 84
RELU
5. Fully Connected	input = 84. Output = 43

3. How model is trained

To train the model, I used an AdamOptimizer, with rate = 0.001, EPOCHS = 30, BATCH_SIZE = 128.

AdamOptimizer is known as computationally efficient, low memory consumption and easy to tune optimizer. Rate

rate and BATCH_SIZE were selected via trial and error method. Smaller rate makes learning too slow, high one does not converge well.

EPOCHS was chosen to achieve the necessary precision.

4. The approach.

My final model results were:

• training set accuracy of 0.999
• validation set accuracy of 0.945
• test set accuracy of 0.933

Some thoughts about the architecture:

• Current LaNet architecture was one I knew from previous course lessons and it did its job.
• For current architecture was quite tricky to find proper hyperparameters and normalize input.
• I tuned the EPOCHS parameter, increased it to 20 to reach learning threshold.

Test a Model on New Images

1. Choose five German traffic signs found on the web and provide them in the report.

Here are five German traffic signs that I found on the web:

2. Result of prediction.

Image	Prediction (top 5 softmax probabilities)
Speed limit (30km/h)	[9.9999928e-01 4.6627065e-07 1.8073200e-07 1.1348734e-08 4.4676702e-12]
General caution	[1.0000000e+00 1.7135681e-08 4.0559920e-09 3.0214271e-09 1.0853046e-09]
Priority road	[1.0000000e+00 7.0696413e-24 1.4433812e-24 1.0934178e-25 5.1488289e-26]
No entry	[1.0000000e+00 5.7582330e-20 2.2768636e-24 8.4445566e-30 7.0737169e-31]
Road work	[1.0000000e+00 3.8328948e-15 6.1613616e-16 3.8493330e-16 7.0987985e-17]
End of all speed and passing limits	[9.9997044e-01 2.9554330e-05 2.5264612e-08 2.8319382e-11 8.3975883e-14]
Stop	[1.0000000e+00 9.8190428e-17 3.0260198e-19 1.7365903e-19 1.4812312e-19]
Keep right	[1.0000000e+00 8.9166329e-25 7.2660056e-32 4.0964350e-32 1.1183832e-32]

The model was able to correctly guess all the signs, which gives an accuracy of 100%. This is the expected result because newly found images are clear, without noise and have good lighting conditions. Taking into account that the model has shown > 93% accuracy on the test set, there should be no problem to classify these images.

Visualizing the Neural Network

Layer	Image
Original image:
Conv1
Conv2