Traffic Sign Recognition

A Simple CNN Architecture inspired from Le-Net 5

---

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. Summary of Data

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

• 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. Visualization of Dataset

A visualization of the frequency of samples per class in the Training dataset can be seen below.

Some samples from the dataset can be seen below.

Design and Test a Model Architecture

1.Data Preprocessing

As a first step, I decided to convert the images to grayscale because it reeduces the number of channels which reduces the training time.

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

I also normalised the pixel values by subtracting 125 and dividing by 125. This restraints them to be in the range of -1 to +1. This makes sure that while backpropogation, the gradients don't blow up.

2. Final Model

My final model consisted of the following layers:

Layer	Description
Input	32x32x1 Grayscale image
Convolution 5x5	1x1 stride, same padding, outputs 28x28x6
RELU
Convolution 5x5	1x1 stride, valid padding, outputs 24x24x12
RELU
Max pooling	2x2 stride, outputs 12x12x12
Convolution 5x5	1x1 stride, valid padding, outputs 8x8x24
RELU
Convolution 5x5	1x1 stride, valid padding, outputs 4x4x24
RELU
Max pooling	2x2 stride, outputs 2x2x24
Flatten
Dense	outputs 240
RELU
Dropout	Keep_prob 0.7
Dense	outputs 240
RELU
Dropout	Keep_prob 0.7
Dense	outputs 120
RELU
Dropout	Keep_prob 0.7
Dense	outputs 43

3. Training the model

To train the model, I used the Adam Optimizer with a learning rate of 0.001 which I ran for a total of 25 epochs with a batch size of 128. As this was a multi-class classification, I used the Cross Entropy function as my loss function.

4. Analysis of the model

My final model results were:

• training set accuracy of 99.0%
• validation set accuracy of 95.1%
• test set accuracy of 92.3% over

Initially I tried the vanilla LeNet-5 but even for a small data, the model was not overfitting which indicated that a more complex network was required. After increasing the number of Convolution and Dense layers, the training accuracy was almost hitting 99% which indicated Overfitting. Hence I used Dropout layers after all the Dense layers with a dropout probability of 0.3. This increased the validation accuracy drastically.

Test a Model on New Images

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

Here are the results of the prediction:

Image	Prediction
Speed Limit (20km/h)	Speed Limit (20km/h)
Wild Animals Crossing	Wild Animals Crossing
Keep Right	Keep Right
No Passing	Children Crossing
Road Work	Road Work
Stop	Stop
Yield	Yield

The model was able to correctly guess 6 of the 7 traffic signs, which gives an accuracy of 85.6%. These images are taken randomly from the internet.

For the fourth image, the model is relatively sure that this is a End of No Passing sign (probability of 0.6), and the image does not contain a stop sign but No Passing sign. The top five soft max probabilities were

Probability	Prediction
.843	Children Crossing
.107	Right-of-way at the next intersection
.003	Vehicles over 3.5 metric tons prohibited
.0005	Slippery road
.0004	No Passing