ashishraste / carnd-behavioralcloning Goto Github PK

In this project a Convolutional Neural Network (CNN) is used to clone and predict a car's steering angles based on images.

Note: This project uses a Unity-based car simulator developed at Udacity Inc. and a dataset of images that is generated from it, none of which are included in this repository.

• scripts : Source containing the CNN model and training/data-loading routines.
• images : Sample images used in this document.
• drive.py : Udacity provided source to control/feed the steering-angles from a given trained-model, along with other parameters like throttle.
• model.h5 : Trained CNN model using Keras.
• video_track1.mp4 : Video recording of the car driving in autonomous mode on Track 1 of the simulator.

1. Collecting data from the simulator that simulates good driving behavior. For this project, the focus was on having the car driving on the center of the lane.

2. Building a convolution neural network in Keras that predicts steering angles from images.

3. Training and validating the model with a training and validation set.

4. Testing that the model successfully drives for one lap without leaving the road.

Training data for predicting the steering-angles was collected using the simulator. Three laps on Track 1 was run in the desired direction. It was noted that the track had more left-turns compared to right ones. To avoid having the model overfitting on left-steering-angles, additional data was collected by driving the car in the opposite direction on the same track for two laps, to have images similar as follows.

Images from the center camera of the car were used to train the model. Sample images from the left, center and right cameras are shown below.

Left camera image	Center camera image	Right camera image

Collected data contains RGB images with dimensions width=160 pixels and height=320 pixels. Most of the preprocessing steps like cropping the images and normalizing them were done as part of building the CNN model, as described in the section below.

For predicting the steering commands based on the images, the CNN architecture from nVidia on End to End Learning for Self Driving Cars, as shown below, is used. Keras is used to build and train the model.

Following preprocessing steps on the images, which were added as part of the model, were carried out.

• Cropping the images : 70 pixels from the top, containing information other than the lane's features and 25 pixels from the bottom having part of the car's body were cropped from the image.

• Normalizing the images : Intensities of RGB pixels of the images were mean-centered around zero with small standard deviation.

• Architecture summary : Below given is the model's architecture details. In total, five convolutional layers with ReLU activations and four fully-connected layers are used. This architecture is inspired from nVidia's model shown above.

• Regularization : L2 regularization and dropout layers didn't seem to improve the model's accuracy, rather decreased it. Hence they aren't included in this architecture.

• Total number of images used during the training phase is 7040.

• Image dataset is split into train and validation set, where 20% of the images are set aside for the validation set.

• Mean Squared Error (MSE) over the predicted steering-angle and actual steering-angle is used as the cost function.

• Adam optimizer is used to optimize the cost function, leaving us to just tune the number of epochs.

• The model is trained over 3 epochs. For each epoch, images are supplied to the network in batches of size 32.

• During the training phase, a prediction accuracy of ~ 99.987 % over the validation set is obtained.

• Following image shows the MSE accuracy over training and validation sets over 3 epochs. It has been noted that the MSE increased for validation set during 2nd epoch which happens when the model is overfitting, but the accuracy during 3rd epoch proved otherwise.

• To test the trained model, the simulator was initialized in autonomous mode. Using the provided drive.py script, the car received steering and throttle commands.

• After managing the two steep turns (left and right) following the bridge, the predicted steering angles couldn't follow the road's center. Additional data was collected around these areas to make the model train predict the steering-angles that'd bring the car back to the road's center.

• Sample recording of the autonomous run on Track 1 is available in this video.