Image Based Heading Control of a Car Using Neural Networks

Abstract

There has been a lot of development going on in the field of self-driving technology these days. The development in this field is due to the breakthrough advancement in deep learning. Tasks that usually require human interaction are executed by training deep neural networks. The Convolutional Neural Networks (CNNs) recognize image features and patterns using different models allowing them to be fruitful. Various factors of what makes a vehicle automated have been addressed in the world of automobiles. In this project we used an end-to-end learning approach that is of training a Convolutional Neural Network with the help of images for a basic adaption of self-driving car. Raw pixels extracted from images obtained by a single front facing camera are mapped directly to the steering commands by this trained CNN. The CNN has an ability to extract information from images such as the patterns and features allowing the automobile to drive autonomously. We used AirSim simulator to generate our dataset along for the testing purposes. Moreover, Unreal Engine has been used to create an unreal environment for generating a dataset through which we can train our CNN We trained our model with different network architectures such as NVIDIA, Alex Net and ResNet 101 with a technique of Transfer Learning.