End-to-End Self-driving and a Dataset Genaration Environment

Abstract

Self-driving systems can be classified into those based on object recognition and motion control methods and those based on End-to-End learning. This research aims to develop a learning model that generates vehicle control signals only from camera images in an End-to-End method. In order to collect training data to be used for deep learning, we constructed a simulation environment where we can drive a vehicle using CARLA driving simulator. We collected as a dataset pairs of images from a camera mounted on the vehicle and human operations. We use Convolutional Neural Networks and Recurrent Neural Networks for learning and compare the models.