Self Driving Car Nano degree Trafic Signal Classifier project using python, Tensorflow 1.3, openv, LetNet5 architecture

Dataset Summary & Visualization: Analyze and summarize the dataset used i.e, detail of Train, Validate and Test data sets. To visualize the dataset.

2.1. Preprocessing: Describes the preprocessing techniques used and why these techniques were chosen.

2.2. Model Architecture: Provide details of the characteristics and qualities of the architecture, including the type of model used, the number of layers, and the size of each layer. Visualizations emphasizing particular qualities of the architecture are encouraged.

2.3. Model Training: Describe how the model was trained by discussing what optimizer was used, batch size, number of epochs and values for hyperparameters.

2.4. Solution Approach: Describe the approach to finding a solution. Accuracy on the validation set is 0.93 or greater.

Get 5 new images of German Traffic signs from Web and use the model to predict them. Calculate the accuracy of the model w.r.t these 5 images. The top five softmax probabilities of the predictions on the captured images to be outputted.

Image Size = 32 x 32 x 3
Number of Training Images = 34799
Number of Validating Images = 4410
Number of Testing Images = 12630

Number of classes in the dataset is 43 which can be inferred by looking for unique values in Image classified/Label values in y_train or y_valid or y_test.

Images for each class is plotted in Step 1 and Number of images in each class is plotted for y_train, y_valid and y_test.

Pre Processing: As each pixel maximum value is 255, (pixel - 128)/ 128 is the only normalization used to pre-process the color image from the dataset of each image size 32 x 32 x 3. The input images are shuffled only once. Basic LeNet architecture is implemented with small changes by adding Dropout and required accuracy is achieved.

Weights used for training are normalized by considering random values from a truncated normal distribution with mean = 0 and sigma = 0.1.

One Hot encoding for labels in y_train, y_valid and y_test is required for softmax is used. Other Hyperparameter considered are

EPOCHS = 25
BATCH_SIZE = 100
Learning Rate = 0.001

Model Architecture:

LeNet-5 architecture is implemented by adding Dropout at certain points as given below and a training accuracy of 0.953 is achieved. Dropout is a regularization technique that helps model avoid overfitting.

CONV2D 32x32x3 -> 28x28x6
relu
Dropout1(0.75)
max_pool -> 14x14x6
CONV2D -> 10x10x16
relu
dropout2(0.75)
max_pool -> 5x5x16
flatten -> 400
FullyConnected -> 120
relu
dropout3(0.75)
FullyConnected -> 84
relu
dropout4(0.75)
FullyConnected -> 43

Optimizer is used for training the model by minimizing the loss. In our training pipeline Adam optimizer is used.

Training Model:

By training the input dataset over this final architecture project required accuracy can be achieved. Graph below shows for considered Hyperparameter and architecture how for each EPOCH the accuracy keeps increasing.

Solution Approach:

With the initial base architecture without proper regularization technique we couldn’t achieve the required accuracy even after 1000 EPOCHs. On each step certain parameter is adjusted to obtain the final architecture as above and with just 25 epoch a high accuracy can be achieved. Different parameter and accuracy achieved is shown in following graphs.

EPOCH 1000
BATCH 256
Learning RATE 0.001
Without Dropout
This is my first trained model Training accuracy Graph!!

EPOCH 1000
BATCH 256
RATE 0.001
DROPOUT1 0.5 [Only one Dropout used]

EPOCH 1000
BATCH 256
RATE 0.001
2 – DROPOUT 1,2 0.5
Max accuracy: 0.932

EPOCH 1000
BATCH 256
RATE_0.001
2 – DROPOUT 1,2 0.5
Max accuracy: 0.897

EPOCH 1000
BATCH 128
RATE 0.001
2 - DROPOUT1,2 0.75
Max Accuracy : 0.937

EPOCH_1000
BATCH_128
RATE_0.001
2 - DROPOUT1,2 0.5
Max Accuracy : 0.899

Untill now Shuffled input were not used. For the first time input shuffling is tried.
EPOCH_500
BATCH_100
RATE_0.001
DROPOUT1 0.75
Max Accuracy 0.93

Shuffled input
EPOCH_500
BATCH_100
RATE_0.001
2- DROPOUT-1,2 0.75
Max Accuracy : 0.948

In last two graphs we could see accuracy is very high right at 1st EPOCH and could reach high accuracy but still testing accuracy could not go above 0.93. After reading through few architectures then came out with used architecture in section Model Architecture with which high accuracy could be achieved in less EPOCH and testing accuracy aswell passed the requirement.

Following five images are chosen for testing the model and it passed with 100% accuracy.

11 – Right-of-way at the next intersection
10 – No passing for vehicles over 3.5 metric tons
29 – Bicycles crossing
41 – End of no passing
02 – Speed limit (50km/h)

Top 5 softmax based on these images are also determined and are as below.