Behavioral Cloning Project

---

Behavioral Cloning Project

The goals / steps of this project are the following:

• Use the simulator to collect data of good driving behavior
• Build, a convolution neural network in Keras that predicts steering angles from images
• Train and validate the model with a training and validation set
• Test that the model successfully drives around track one without leaving the road
• Summarize the results with a written report

Rubric Points

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

---

Model Architecture and Training Strategy

1. Final Model Architecture

I decided to use the nVidia model architecture, which was proven to give quite good results and consists of 5 convolutional layers, followed by 3 fully connected layers. The model implementation nVidiaModel() can be found in lines #109 through #124 in model.py.

Source: https://devblogs.nvidia.com/parallelforall/deep-learning-self-driving-cars/

I added a pre-processing layer to this model for normalization of the input data. Additionally, I cropped the input images to remove for the driving purpose unnecessary information from the images. I removed the upper 50 % of the image (trees, sky, landscape, far-distance parts of the road) and the lower 20% of the image (parts of the car). This is implemented with createPreProcessingLayers() in lines #100 through #107 in model.py.

2. Creation of the Training Set & Training Process

To capture good driving behavior, I made two recording driving track 1 in forward direction. To generalize better, I made an additional recording driving track 1 backwards. For smoother driving behaviour I used the mouse pointer for steering.

From the beginning, I used data from all three cameras and corrected the steering measurements from the left and right camera images, which is implemente with correctMeasurements() in lines #50 through #64 in model.py. The first try of the model with a correction factor of 0.1 was not successfull. In the end, 0.2 appeared to be a reasonable value.

The second try with 0.2 already gave quite good performance. However, the sharp left turn after the bridge was missed. I made a short recording only with that particular left turn and added that to the training data, which improved the performance and lead to a successful run of the whole track.

Moreover, shuffling and flipping the images appeared for my implementation to be sufficient for data augmentation. 3 epoches were enough for sufficient performance on track 1.

The complete video can be found here: