SDC-TrafficSignDetection

Self-driving car project: Detecting Traffic Signs

Self-Driving Car Project: Traffic Sign Classification

Hello there! I'm Babak. Let me introduce you to my second project in the self-driving car projects series. In this project I built a classifier to detect traffic signs on the road using convolutional nueral network with LaNet architecture, this is an important step in developing a self-driving car. Algorithm for this project is written in Python using TensorFlow, numpy, pickle and matplotlib libraries. The algorithm was trained on german traffic sign dataset and tested on actual traffic sign images.

Contents

• Python Script
• Jupyter notebook
• Test_traffic_signs folder
• Readme file
• Readme_images folder

Pipeline:

The pipeline for detecting lane lines on the road is as follows:

1. Loading the data set (German Traffic Signs.)
2. Exploring, summarizing and visualizing the data set.
3. Preprocessing the data (grayscale, normalize)
4. Designing and implementing a convolutional nueral network model with LeNet-5 architecture.
5. Training, validating and testing the model.
6. Testing the model on new traffic sign images.
7. Analyze the softmax probabilities of the new images

Writeup:

Data Set Summary & Exploration

I used the numpy, pickle and pandas libraries to calculate summary statistics of the traffic signs data set:

• The size of training set: 34799
• The size of the validation set: 4410
• The size of test set: 12630
• The shape of a traffic sign image: 32px x 32px x 3
• The number of unique classes/labels in the data set: 43

Sample of the data set is shown here:

Here is an exploratory visualization of the data set. It is a bar chart showing number of data points per class for training, validation and testing data set.

Data Set Preprocessing

I preprocessed the data. As a first step, I converted the images to grayscale to train efficeintly and less complexity. It is better to train the model on single channel images first.

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

As a last step, I normalized the image data to change the range of pixel intensity to a range that is more normal to the senses. The mean of the normalized image is 0 and variance is 1.

Here is an example of an original image and a normalized image:

Design and Test a Model Architecture

My final model architecture wass LeNet-5 and it consisted of the following layers:

Layer	Description
Convolution 5x5	input: 32x32x1, output: 28x28x6
RELU	Activation layer
Max pooling	input: 28x28x6, output: 14x14x6
Convolution 5x5	input: 14x14x6, output: 10x10x16
RELU	Activation layer
Max pooling	input: 10x10x16, output: 5x5x16
Flattening	input: 5x5x16, output: 400
Fully connected	input: 400, output: 120
RELU	Activation layer
Dropout	Dropout Layer
Fully connected	input: 120, output: 84
RELU	Activation layer
Dropout	Dropout Layer
Fully connected	input: 84, output: 43

To train the model, I used the Adam optimizer. The final hyper parameters used were: batch size: 120 epochs: 100 learning rate: 0.0007 mu: 0 sigma: 0.1 dropout keep probability: 0.5

My final model results were:

• validation set accuracy: 97.3%
• test set accuracy: 95%

###Test a Model on New Images

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

The images I chose were very difficult and challenging to classify because they're at different angles, distances with various backgrounds. Also, the German Traffic Sign dataset contain a border of 10 % around the actual traffic sign to allow for edge-based approaches but my images don't have such a border. This could be another source of confusion for the model. I applied the preprocessing i.e. grayscaling and normalizing to my new images.

Here are the results of the prediction:

[20, 12, 25, 17, 35] <-predictions labels

[4, 38, 16, 17, 25] <-actual labels

Test Set Accuracy = 0.200

As can be seen from the labels results above and also the softmax probabilities below, the newly acquired images did poorly compared to the previously obtained ones. The reason is that I intentionally chose challenging and difficult German traffic signs to really test my model and see the response for noisy, angled, zoomed with various background images.

The model was able to correctly guess 1 of the 5 traffic signs, which gives an accuracy of 20%.

The top soft max probabilities for each image were: