Autonomous Driving by Using Convolutional Neural Network

Autonomous driving is one of the research areas that are in the main focus nowadays, because it can improve humans' life and safety in driving. The purpose of this research work is to train a neural network that will imitate a driver's driving. This will be based on a data set from his previous drivings. We use some of the most used libraries today related to work with neural networks such as TensorFlow and Keras. Our model and the architecture of the neural network are based on a research paper from the company NVIDIA, where they trained a neural network that maps pixels from a front positioned camera into direct steering wheel commands on the actual car. In our research we don't use actual car but rather a free to use online simulator that Udacity is providing. This simulator has 3 front positioned cameras, which save our driving around a circuit. The actual data that is recorded consist of 3 pictures from the 3 front mounted cameras (left, central and right camera), while also saving the steering wheels angle at a specific time stamp. With this simulator we collect the data. At the end, once the model is trained well enough, we test it on the “autonomous” mode that the simulator is offering, where our vehicle will try to drive autonomously around the track.