Publication: Técnicas inteligentes para su integración en un vehículo autómata
Loading...
Official URL
Full text at PDC
Publication Date
2020
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Es evidente que los vehículos autónomos de futuro constituyen un nicho de interés tecnológico elevado. La dotación de inteligencia a los mismos es, a su vez, un reto. Es bajo esta perspectiva sobre la que se plantea el presente trabajo, si bien a muy pequeña escala.
En efecto se propone el desarrollo de un prototipo de vehículo autónomo dotado de inteligencia, esto es, un autómata inteligente. El modelo de prototipo, en línea con desarrollos de este tipo, se plantea considerando una plataforma dotada con los elementos necesarios para navegación autónoma, incluyendo ruedas, motores, alimentación, por supuesto sensores de visión (cámara) y de proximidad (ultrasonidos) y naturalmente un computador local basado en una Raspberry Pi que gestiona y controla las acciones del vehículo. El sistema también consta de un computador remoto, en perfecta comunicación y sincronía con el local donde se realizan los procesos con alta carga computacional y donde está instalada la parte inteligente del sistema.
El computador local envía datos (imágenes y distancias) al remoto, iniciándose en éste su procesamiento. En el caso de las imágenes mediante la aplicación de técnicas de segmentación para localizar una plataforma con una base rectangular negra conteniendo figuras geométricas (cuadrado, círculo, rectángulo, triángulo) de color en su interior, y que constituye el objetivo del vehículo hacia el que se tiene que dirigir. Con tal finalidad, se extraen regiones candidatas, que pueden contener la plataforma, aplicando técnicas de umbralización y etiquetado de componentes conexas, para proceder a recortar dichas regiones sobre la imagen original. Estos recortes se pasan a una Red Neuronal Convolucional, previamente re-entrenada a partir del modelo AlexNet, que proporciona un valor de probabilidad para determinar si el recorte puede o no considerarse como la plataforma. Un proceso posterior, basado en la identificación de las figuras geométricas de la plataforma refuerza o penaliza la probabilidad dada por la red, procediendo a decidir si la región analizada es o no finalmente la plataforma. En el caso de las distancias, proporcionadas por el sensor de ultrasonidos, simplemente se trata de verificar la proximidad del vehículo a un objeto o a la plataforma, para confirmar la llegada al objetivo. Esta información se transmite a la Raspberry Pi, que se convierte en acciones sobre el vehículo para girar, avanzar y seguir o no enviando más datos desde el vehículo.
Los experimentos realizados han verificado la validez de la propuesta, destacando los aspectos relacionados con el sistema inteligente y la navegación autónoma del vehículo autómata.
It is obvious the autonomous vehicles of the future constitute a niche of high technological interest. The endowment of intelligence to them is, in turn, a challenge. It is from this perspective that the present work is proposed, albeit on a very small scale. Indeed, the development of a prototype of an autonomous vehicle endowed with intelligence is proposed, i.e. an intelligent automaton. The prototype model, in line with developments of this type, is proposed considering a platform equipped with the necessary elements for autonomous navigation, including wheels, motors, power supply, of course vision-based sensors (camera) and proximity (ultrasound) and obviously a local computer based on a Raspberry Pi, which handles and controls the actions of the vehicle. The system also consists of a remote computer, in perfect communication and synchrony with the local computer, where processes with high computational load are carried out and where the intelligent part of the system is installed. The local computer sends data (images and distances) to the remote, starting its processing there. In the case of images by applying segmentation techniques to locate a platform with a black rectangular base containing geometric colored figures (square, circle, rectangle and triangle) inside, and which constitutes the objective of the vehicle towards it has to be guided. For this purpose, candidate regions, which may contain the platform, are extracted, by applying thresholding techniques and labeling of connected components, in order to clip these regions on the original image. These clipping are passed to a Convolutional Neural Network, previously re-trained from the AlexNet model, which provides a probability value to determine whether or not the clipping can be considered as the platform. A subsequent process, based on the identification of the geometric figures of the platform, reinforces or penalizes the probability given by the network, proceeding to decide whether or not the analyzed region is finally the platform. In the case of distances, provided by the ultrasound sensor, it is simply a matter of verifying the proximity of the vehicle to an object or to the platform, to confirm arrival at the target. This information is transmitted to the Raspberry Pi, which is converted into actions on the vehicle to turn, advance and continue or not sending more data. The experiments carried out have verified the validity of the proposal, highlighting aspects related to the intelligent system and autonomous navigation of the automated vehicle.
It is obvious the autonomous vehicles of the future constitute a niche of high technological interest. The endowment of intelligence to them is, in turn, a challenge. It is from this perspective that the present work is proposed, albeit on a very small scale. Indeed, the development of a prototype of an autonomous vehicle endowed with intelligence is proposed, i.e. an intelligent automaton. The prototype model, in line with developments of this type, is proposed considering a platform equipped with the necessary elements for autonomous navigation, including wheels, motors, power supply, of course vision-based sensors (camera) and proximity (ultrasound) and obviously a local computer based on a Raspberry Pi, which handles and controls the actions of the vehicle. The system also consists of a remote computer, in perfect communication and synchrony with the local computer, where processes with high computational load are carried out and where the intelligent part of the system is installed. The local computer sends data (images and distances) to the remote, starting its processing there. In the case of images by applying segmentation techniques to locate a platform with a black rectangular base containing geometric colored figures (square, circle, rectangle and triangle) inside, and which constitutes the objective of the vehicle towards it has to be guided. For this purpose, candidate regions, which may contain the platform, are extracted, by applying thresholding techniques and labeling of connected components, in order to clip these regions on the original image. These clipping are passed to a Convolutional Neural Network, previously re-trained from the AlexNet model, which provides a probability value to determine whether or not the clipping can be considered as the platform. A subsequent process, based on the identification of the geometric figures of the platform, reinforces or penalizes the probability given by the network, proceeding to decide whether or not the analyzed region is finally the platform. In the case of distances, provided by the ultrasound sensor, it is simply a matter of verifying the proximity of the vehicle to an object or to the platform, to confirm arrival at the target. This information is transmitted to the Raspberry Pi, which is converted into actions on the vehicle to turn, advance and continue or not sending more data. The experiments carried out have verified the validity of the proposal, highlighting aspects related to the intelligent system and autonomous navigation of the automated vehicle.
Description
Trabajo de Fin de Grado en Ingeniería del Software, Facultad de Informática UCM, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2019/2020.