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

Rubric Points

Here I will consider the rubric points individually and describe how I addressed each point in my implementation.

---

You're reading it! and here is a link to my project code

Data Set Summary & Exploration

I used the numpy 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

Here is an exploratory visualization of the data set. It is a bar chart showing the distribution of the different signs.

Pre-Processing

At first I tried to use one of the normaalization methods in the provided in the course content: (X_train - 128) / 128. I tried this but actually gave me worse results. Instead in my network I use tensorflows very own per_image_standarization. There was no need to resize or crop the image because this had already been done by the provider of the dataset.

Design and Test a Model Architecture

I started out with the LeNet architecture we learned in a preview exercise. This offered about 89% validation accuracy.

Next, I played with learning rete and number of epocs. When lowering my learning rate and raising epocs it seemed bump my validation accuracy into the lower 90s. In the end this didn't seem to be a big factor. I ended up resetting my learning rate to .001 and bumping the epocs from the default LeNet value of 10 to 20.

Next, I tried playing with the regularization to prevent overfitting. LetNet already had some older tactics with MaxPooling. First, I added a dropout in my last fully connected layer. This bumped my validation accuracy up to the 93 range. When I tried added dropout to the other layers including conv layers with .2 keep_prob accuracy seemed to decrease. So I removed those.

Lastly, I decided to add another convolution to the network. This bumped my accuracy into the 95-96 range.

Also, as stated above:

Knowing that the images at a minimum needed to be normalized I used a standard normalization tensor from tensorlfow: per_image_standardization. I did this right in the network as stated above.

My final model consisted of the following layers:

Layer	Description
Input	32x32x3 RGB image
Convolution 3x3	1x1 stride, valid padding, relu, 2x2 maxpool, output 32x32x32
Convolution 3x3	1x1 stride, valid padding, relu, 2x2 maxpool, output 32x32x64
Convolution 3x3	1x1 stride, valid padding, relu, 2x2 maxpool, output 32x32x128
Fully Connected	flattened
Fully Connected	relu, outputs 512X120
Fully Connected	relu, outputs 120X84
Fully Connected	relu, dropout, outputs 84x43

Training:

To train the model, I used a batch size of 128 and 20 epocs. Used sklearn's shuffle to get random mini batches. Played with the learning rate but landed on a go to learning rate of .001. For my dropout function I stuck with .5 keep_prob. To monitor the training process the training batch and validation batch are evaluated for accuracy. When all 20 epochs are completed we save the model to evaluate the test and internet images in the upcoming cells.

Iterative Approach

I made many tweaks to get to the final network. First I was playing with the LeNet. Tweaks to learning rate, dropout and then extra layers and filter depth.

My final model results were:

• training set accuracy of 100.0%
• validation set accuracy of 96.0%
• test set accuracy of 94.6%

Test a Model on New Images

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

The images I selected were pretty clear but some of them share some of the same shapes of other images in different classes.

Here are the results of the prediction:

Image	Prediction
Road Work	Road Work
Speed 30	Speed 30
Yield	Yield
Left Turn	Left Turn
right_of_way_next_intersection	right_of_way_next_intersection

The model was able to correctly guess 5 of the 5 traffic signs, which gives an accuracy of 100%. This compares favorably to the accuracy on the test set of 94.6% considering the quaility of the images and the small sample size.

All of my internet imaages were almost 100% what the answer was. In my IPython notebook there is dataframes for each of the images. I'll show the worst performing one here:

Probability	Prediction
99.90%	Road work
0.08%	Children crossing
0.01%	End of all speed and passing limits
0.01%	Pedestrians
0.00%	Beware of ice/snow