Desarrollo de un coche a escala reducida para aprendizaje de conducción autónoma mediante técnicas de Deep Learning
Year: 2020
Type: dissertation
Abstract: Technological advances in autonomous driving are growing at a vertiginous rate and \nare gradually being integrated into society. The possibility of reducing the number of \naccidents, reducing the effort required to drive and facilitating travel for the disabled are \nsome of the advantages of autonomous vehicles. \nIn this project, a scale vehicle will be developed that learns to drive autonomously. \nThis will require the combination of a series of important elements, such as the scale \nvehicle, a driving circuit and a computer with a control device (steering wheel). This \nsetting should allow both supervised driving and autonomous driving learning. \nTo perform autonomous driving training, an approach divided into two tasks is \nproposed: extracting features from road lines and learning to drive from these features. \nFor the first task, the construction of an auxiliary scenario (based on chromas) is proposed \nto facilitate the extraction of the lines. These lines can be virtually superimposed on \nimages of the driving scenario, forming augmented images that will serve to train an \nAutoencoder. In this way, this network will learn to isolate the lines from the road in real \nscenarios, making it possible to learn to drive in them. \nAs a result, a small-scale vehicle is obtained that is capable of driving autonomously \nalong the route of a circuit delimited by two lines of road. It has been observed how the \naddition of noise during the learning process can improve the robustness of a neural \nnetwork. The auxiliary scenario has made possible to differentiate the road lines from the \nrest of the elements of the scene. In this way it has been possible to obtain a vector of \nfeatures related only to these lines, to generate augmented images of a real scenario and \nto learn to drive on it.
Author: Corrochano Jiménez, Javier
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6990308269