IoT para monitorización de tránsito peatonal mediante técnicas de aprendizaje profundo
Loading...
Official URL
Full text at PDC
Publication date
2019
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
En este trabajo se aborda el diseño y desarrollo de un sistema centrado en el cómputo del tránsito peatonal a partir de secuencias de vídeo grabadas en un determinado escenario. Se propone una solución conceptual basada en el aprovechamiento de tecnologías IoT. El cómputo de dicho tránsito puede dividirse en dos partes: primeramente se toman las imágenes de la secuencia de vídeo grabada, y en cada una de ellas se aplican una serie de técnicas de procesamiento de vídeo y detección de movimiento. La aplicación de estas técnicas da como resultado una serie de recortes para cada una de las imágenes, estos recortes son zonas donde se ha identificado que hay algún peatón presente. La segunda parte consiste en el uso de una Red Neuronal Convolucional para llevar a cabo la clasificación de los recortes obtenidos previamente. Los recortes son clasificados por la red neuronal en cuatro categorías diferentes (unCaminante, dosCaminantes, corredor y ciclista). La clasificación permite obtener los datos necesarios para llevar a cabo el cómputo del tránsito de peatones en el escenario en el que se grabó la secuencia. Los datos obtenidos son subidos a una plataforma en la nube, lo que permite que sean descargados por parte de la administración o de cualquier persona que esté interesada en dichos datos, en ambos casos suponiendo que se tiene acceso, principalmente mediante internet. Por otro lado, en este trabajo también se ha realizado el estudio y predicción a partir de series temporales. Para ello se han aplicado dos modelos diferentes: un modelo autorregresivo y el uso de redes Long Short-Term Memory (LSTM).
This work addresses the design and development of a system focused on pedestrian traffic accounting from video sequences recorded in a given scenario. The paper proposes a conceptual solution based on using IoT technologies. Traffic accounting can be divided into two parts: first, the images of the recorded video sequence are captured, and in each of them a series of video processing and motion detection techniques are applied. The application of these techniques results in a series of clippings for each image; these clippings are areas where it has been identified that a pedestrian is present. The second part consists of the use of a Convolutional Neural Network to carry out the classification of the previously obtained clippings. The clippings are classified by the neural network in four different categories (oneWalker, twoWalkers, runner, and cyclist). The classification allows obtaining the necessary data to carry out the pedestrian accounting in the scenario in which the sequence was recorded. The classification allows obtaining the necessary data to carry out the pedestrian accounting in the scenario in which the sequence was recorded. The obtained data is uploaded to a cloud platform, which allows it to be downloaded by the administration or any person who is interested in such data, assuming in both cases that they have access, mainly through internet. On the other hand, in this paper, the study and prediction of data from time series have also been carried out. For this, two different models have been applied: an autoregressive model and the use of Long Short-Term Memory (LSTM) networks.
This work addresses the design and development of a system focused on pedestrian traffic accounting from video sequences recorded in a given scenario. The paper proposes a conceptual solution based on using IoT technologies. Traffic accounting can be divided into two parts: first, the images of the recorded video sequence are captured, and in each of them a series of video processing and motion detection techniques are applied. The application of these techniques results in a series of clippings for each image; these clippings are areas where it has been identified that a pedestrian is present. The second part consists of the use of a Convolutional Neural Network to carry out the classification of the previously obtained clippings. The clippings are classified by the neural network in four different categories (oneWalker, twoWalkers, runner, and cyclist). The classification allows obtaining the necessary data to carry out the pedestrian accounting in the scenario in which the sequence was recorded. The classification allows obtaining the necessary data to carry out the pedestrian accounting in the scenario in which the sequence was recorded. The obtained data is uploaded to a cloud platform, which allows it to be downloaded by the administration or any person who is interested in such data, assuming in both cases that they have access, mainly through internet. On the other hand, in this paper, the study and prediction of data from time series have also been carried out. For this, two different models have been applied: an autoregressive model and the use of Long Short-Term Memory (LSTM) networks.
Description
Trabajo de Fin de Máster, Universidad Complutense, Facultad de Informática, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2018/2019