Sistema de Localización para Vehículos Autónomos Basado en Visión Artificial

Abstract

The present thesis aims to develop a system for the georeferencing of an autonomous vehicle based on computer vision, with a specific focus on the use of neural networks, applied on the Amazon Web Services (AWS) DeepRacer platform.

An innovative application of Convolutional Neural Network (CNN) as a regressor was explored to address the challenging task of localization in the AWS DeepRacer autonomous vehicle. The CNN, trained with a comprehensive dataset of images and spatial coordinates, proved effective in accurately predicting locations in a two-dimensional environment. Additionally, the perfor mance of the CNN was thoroughly evaluated in comparison to the Lighthouse Position Deck positioning system, highlighting the latter’s importance as a crucial reference framework. Spe cific CNN models that exhibited outstanding performance, minimizing discrepancies between predictions and actual coordinates, were identified. The research provided valuable insights into the effectiveness of CNNs in specific environments, establishing a foundation for the continuous optimization of localization systems in AWS DeepRacer. The comparison between CNN and the Lighthouse Position Deck system contributed to a comprehensive understanding of the approa ches, their strengths and limitations, and pointed out areas for future research in the field of autonomous vehicles. The application of computer vision and neural networks plays a central role in this project, empowering the autonomous vehicle to process and interpret its environment through visual data. These technologies enable the identification of landmarks and local features, facilitating the precise calculation of the vehicle’s position in its surroundings. This approach represents a prominent application of artificial intelligence and deep learning in the context of autonomous navigation, emphasizing the relevance and impact of these innovations in the development of autonomous vehicles.