Behavioral Cloning

Project No3 for 1st. Term of Self-Driving Car Engineer Nanodegree at Udacity

---

Behavioral Cloning Project

The steps of this project are the following:

• Use the simulator to collect data of good driving behavior. To achieve this task, Most of the data that I used is from Udacity's sample training data, I collected some data from my driving and I performed image augmentation.
• Build, a convolution neural network in Keras that predicts steering angles from images. I used NVIDIA MODEL
• Train and validate the model with a training and validation set.
• Test that the model successfully drives around track one without leaving the road

Rubric Points

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

---

Files Submitted & Code Quality

My project includes the following files:

• model.py containing the script to create and train the model (also the Model.ipynb file contains the details about the process)
• drive.py for driving the car in autonomous mode
• model.h5 containing a trained convolution neural network
• readme.md summarizing the results
• Track1.mp4
• Track2.mp4

Model Architecture and Training Strategy

1. Collecting the Data

I only used only Udacity's sample training data and image augmentation.

Images captured from the Udacity simulator:

• [Download Linux Simulator] (https://d17h27t6h515a5.cloudfront.net/topher/2017/February/58983558_beta-simulator-linux/beta-simulator-linux.zip)
• [Download MacOS Simulator] (https://d17h27t6h515a5.cloudfront.net/topher/2017/February/58983385_beta-simulator-mac/beta-simulator-mac.zip)
• Download Windows Simulator

a.Image processing:

All this process have been implemented in the read_image function.

i. Select the data that can represent a good data distribution (removing some images with steering =0 ), and eliminate some data that has an unexpected behavior (Sometimes I got out of the road). This is my data distribution before image augmentation:

ii. Select area of interest, the dimensions of the image is 320x160 pixels. The area of interes is only the road, then I cropped the image to get only the road.

iii. Resize the image to 200x66 pixels and generate a numpy array of float values img =np.asarray(img, dtype=np.float32) This will be the data entry for the convolutional neuronal network.

It is returning the following image:

b. Image augmentation consist in the following functions:

All this process have been implemented in the preprocess_image function, and for the the batch generation I'm using generate_train_batch function

i. Flip the image to have more steering angles to analyze. It involves to multiply the steering*-1

ii. Use right and left cameras. The steering angle would be less than the steering angle from the center camera. From the right camera's perspective, the steering angle would be larger than the angle from the center camera. steering =steering +/- delta I defined this delta as 0.25 (This data was obtained after evaluation on the pixels of images)

iii. Change brightness in order to get a more rebust model to the brightness.

iv. Image translation to generate more images with steering variations.

v. Little perturbation in the steering, this because the the same image could have a little variation on the steering that will be a good answer, even for the humans we perform little different variations of the steering in the same situation. nsteering= steering*(1+np.random.uniform(-1, 1)/50)

2. NVIDIA Model .

I used NVIDIA MODEL : My model consists has as an input a RGB (3 filters) image with 200x66 as dimensions, then the input is (200,66,3 )

*The data is normalized in the model using a Keras lambda layer (code in the function get_NVIDIA_model). This is useful to have a more familiar range to work on. The reason is because in the process that we will train our network.

• I used ELU (Exponential Linear Units) as activation function. The “exponential linear unit” (ELU) which speeds up learning in deep neural networks and leads to higher classification accuracies. [Reference] (https://arxiv.org/pdf/1511.07289v1.pdf)
• I applied Dropout to the input (code in function get_NVIDIA_model). Dropout consists in randomly setting a fraction p of input units to 0 at each update during training time, which helps prevent overfitting. in this case p=0.5
• I used Adam optimizer with the default parameters `keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)' link
• In order to gauge how well the model was working, I split my image and steering angle data into a training and validation set. The validation set is 1/8 of the training set.
• For Training and Validation Loss Metrics, I used fit_generator because it runs in parallel to the model, for efficiency.
• The model was trained and validated on Udacity's sample training data with Image augmentation to ensure that the model was not overfitting. The model was tested by running it through the simulator and ensuring that the vehicle could stay on the track. However, I have to modify the drive behavior to get this result. Here is a visualization of the architecture:

Image generated with plot(model, to_file='model.png',show_shapes =True, show_layer_names=False)

Here is a visualization of the mean squared error loss in the final model elected: