vigneshr97 / traffic-sign-classifier Goto Github PK

In this project, traffic signs are classified by training and validating a model using the German Traffic Sign Dataset. After the model is trained, German traffic signs that are found on the web are classified using the model

The goals / steps of this project are the following:

• Load the 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

The function unique() found in numpy library was used to calculate the size of the three sets. The sizes of Training, Validation and Test sets were 34799, 4410, 12630 respectively from 43 different categories of traffic signs. Each image was 32px by 32px.

Here is a bar chart of the training data set

Here is a bar chart of the validation data set

Here is a bar chart of the test data set

As a first step, I decided to convert the images to grayscale. Then the images were normalized for every value to lie between -1 and 1.

I decided to generate additional data because the test accuracy values were really low and amount of data provided is low. After data augmentation, the number of images in each category was 2500 which was almost thrice the original size. The data augmentation was performed using various image processing techniques. The techniques used were median blurring, Gaussian blurring, Histogram Equalization (CLAHE), Brightness modification, random rotation of upto 15 degrees in either direction and translation of upto 5 pixels on either sides.

The difference between the original data set and the augmented data set is the following ...

LeNet architecture was chosen and a few modifications were made:

• Since LeNet was developed for Character recognition which is almost similar to a traffic sign classification problem, the same architecture was chosen with a few modifications
• Another fully connected layer with RELU was added and dropout was introduced in both the layers to prevent overfitting
• An extra fully connected layer was added and it was observed to perform better.
• Adam optimizer was chosen as it performed better in previous experiences (including behavioral cloning project). The Stochastic Gradient descent is believed to improve slower than the Adam optimizer

My final model consisted of the following layers:

To train the model, I used an Adam optimizer with a tunable learning rate which was set to 0.001. The number of epochs was 50 initially and it was still improving. Each run of the experiment had a different stopping point for improvement. So a safer higher epoch value of 100 was set. The batch size was set to 256 because the accuracy was almost similar to when it was set to 64 and 128 after a few epochs. The weights were initialized from a truncated normal distribution with mean 0 and a standard deviation of 0.06. Those values were chosen by testing the accuracy by changing both the values. The model improves much faster when these two values were used.

My final model results were:

• training set accuracy of 99.2%
• validation set accuracy of 96.9%
• test set accuracy of 94.9%

LeNet architecture was chosen and a few modifications were made:

• Since LeNet was developed for Character recognition which is almost similar to a traffic sign classification problem, the same architecture was chosen with a few modifications
• Another fully connected layer with RELU was added and dropout was introduced in both the layers to prevent overfitting
• Adam optimizer was chosen as it performed better in previous experiences (including behavioral cloning project). The Stochastic Gradient descent is believed to improve slower than the Adam optimizer
• The parameter tuning started with changing mean and standard deviation of the normal distribution from where the values of weights are chosen. Then learning rate, batch size and epochs are tuned in the same order. Then dropouts were introduced and a significant improvement was observed

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

The first image might be difficult to classify because of its similarity with 80 and its not a classifier focussed on numbers. So distinctive features of each number were not taken into account unlike a number classifier. After running it on more epochs, this problem got resolved. The second image gave its result as a bumpy road a few times.

Here are the results of the prediction:

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 94.9%

The code for making predictions on my final model is located in the cells 13-17 of the Ipython notebook.

For the first image, the model is very sure that the sign is a speed limit of 60 Kmph (probability of 1.0), and the image does contain a 60 Kmph speed limit. The top five soft max probabilities were (The other probabilities were less than 0.001. So they aren't listed here. The top 5 probabilities all have number related signs

For the second image, the model is almost sure with a probability of 0.868 with a bumpy road being predicted with a probability of 0.132. The other probabilities were almost zero with the order being Children crossing, Dangerous curve to the right and slippery road. In many other runs, the bumpy road sign was the one that was actually predicted.

The last three signs encountered precise and accurate predictions with the probabilities of predictions being almost equal to 1. Predictions for third image

Predictions for fourth image

Predictions for fifth image