Autonomous driving has been an active area of research for many years, especially since the rise of deep learning for computer vision tasks. In the past several years, many different works used end-to-end learning in an attempt to deal with the complexity of the problem, relying on datasets of collected driving data to train their models.
This thesis presents an overview of some of these papers and highlights decisions, considerations and difficulties encountered when creating an autonomous driving model using end-to-end learning. A convolutional neural network is trained to predict steering wheel angles using a dataset of front view images of human driving collected by AutoNOMOS Labs. Different configutations and settings for dataset composition and network input and output are explored and compared. The models are evaluated on a validation subset and with the use of the VisiualBackProp algorithm, which visualizes regions of the input with the biggest influence on the network prediction.
The best performing model is able to follow the lane on highways and is shown to have learned to detect and recognize lane markings to be used in the decision-making process.