Publication: Reconocimiento de vehículos y su movimiento mediante redes neuronales convolucionales en vídeos
Loading...
Official URL
Full text at PDC
Publication Date
2022
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Este trabajo tiene como finalidad desarrollar una aplicación, que utiliza técnicas de Aprendizaje Profundo, o Deep Learning (DL), en Inteligencia Artificial para la detección de movimiento de vehículos en secuencias de imágenes, con fines de su identificación dentro del flujo de tráfico y como posible solución en las futuras ciudades inteligentes.
El objetivo es proporcionar una solución conceptual para el control y seguimiento de vehículos sobre la vía urbana. Dentro del AP se utilizan concretamente dos modelos de Redes Neuronales Convolucionales (RNC), exactamente AlexNet y GoogleNet para la identificación de los vehículos en movimiento, el cual se detecta mediante técnicas basadas en flujo óptico como técnica propia de Visión por Computador. Estos modelos requieren de un entrenamiento previo. Así se consigue extraer rasgos característicos de los vehículos, a partir de una serie de imágenes secuenciadas en vídeos, que pueden ser previamente seleccionados por el usuario.
La aplicación desarrollada en este proyecto permite al usuario modificar los parámetros necesarios para el entrenamiento de las RNC, con la finalidad de poder obtener un resultado más eficiente y óptimo. Tras realizar el entrenamiento, el usuario también puede seleccionar previamente un video para generar la clasificación de los vehículos e identificarlos dentro del tráfico de vehículos.
Complementariamente, los datos obtenidos tras la clasificación llevada a cabo se subirán a la nube mediante el uso de la plataforma ThingSpeak, como servicio proporcionado por Matlab dentro del paradigma de Internet de las Cosas (IoT, Internet of Things) que permite el almacenamiento y análisis de distintas categorías de datos. Esta plataforma permite mostrar los resultados y enviar mensajes a través de la red social Twitter para su posterior valoración e interpretación.
The purpose of this work is to develop an application that uses Deep Learning techniques in Artificial Intelligence for vehicle movement detection in image sequences, with the purpose of identifying them within the traffic flow and as a possible solution in future smart cities. The objective is to provide a conceptual solution for the control and monitoring of vehicles on urban roads. Within the Deep Learning (DL), two models of Convolutional Neural Networks (CNN) are specifically used, AlexNet and GoogleNet, for the identification of moving vehicles which is detected using techniques based on optical flow, as a Computer Vision technique. These models require previous training. It is possible to extract typical features of the vehicles, from a series of images sequenced in videos, which can be previously selected by the user. The application developed in this project allows the user to modify the necessary parameters for the CNN training, in order to obtain a more efficient and optimal result. After completing the training. the user can also previously select a video to generate the vehicle classification and identify them within the vehicle traffic. In addition, the data obtained after the classification carried out will be uploaded to the cloud by using the ThingSpeak platform, as a service provided by Matlab within the Internet of Things paradigm that allows storage and analysis of different categories of data. This platform allows the results to be displayed and sent via the social network Twitter for subsequent assessment and interpretation.
The purpose of this work is to develop an application that uses Deep Learning techniques in Artificial Intelligence for vehicle movement detection in image sequences, with the purpose of identifying them within the traffic flow and as a possible solution in future smart cities. The objective is to provide a conceptual solution for the control and monitoring of vehicles on urban roads. Within the Deep Learning (DL), two models of Convolutional Neural Networks (CNN) are specifically used, AlexNet and GoogleNet, for the identification of moving vehicles which is detected using techniques based on optical flow, as a Computer Vision technique. These models require previous training. It is possible to extract typical features of the vehicles, from a series of images sequenced in videos, which can be previously selected by the user. The application developed in this project allows the user to modify the necessary parameters for the CNN training, in order to obtain a more efficient and optimal result. After completing the training. the user can also previously select a video to generate the vehicle classification and identify them within the vehicle traffic. In addition, the data obtained after the classification carried out will be uploaded to the cloud by using the ThingSpeak platform, as a service provided by Matlab within the Internet of Things paradigm that allows storage and analysis of different categories of data. This platform allows the results to be displayed and sent via the social network Twitter for subsequent assessment and interpretation.
Description
Trabajo de Fin de Grado en Ingeniería del Software, Facultad de Informática UCM, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2021/2022.