Publication: Técnicas de Deep Learning para el reconocimiento de movimientos corporales
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
En este trabajo se detalla el diseño y desarrollo de un sistema de reconocimiento de actividades humanas (HAR, human activity recognition) mediante redes neuronales, así como la labor de investigación realizada sobre técnicas y proyectos similares realizados sobre este mismo campo. El sistema desarrollado consta de varios módulos que hacen uso e interactúan a través de internet, estando en parte alojados en la nube.
Concretamente, el diseño de la aplicación contiene dos módulos principales, una aplicación móvil para sistemas Android y una aplicación web alojada en un servidor. La primera, instalada en un dispositivo móvil, recaba datos de los sensores del dispositivo (sensores de aceleración y movimiento) y los envía al servidor por internet para ser analizados. La aplicación web consiste en una api REST, que hace uso de una red neuronal entrenada para realizar predicciones sobre la actividad que se está realizando, basándose en los datos recibidos. El tipo de red empleado es una red recurrente de tipo Long-Short-Term Memory (LSTM), basada en los trabajos publicados por Matlab sobre clasificación y predicción sequence-to-sequence. Esta herramienta es también la que se ha empleado para desarrollar el modelo de red y realizar los entrenamientos necesarios.
El proceso de predicción se inicia cuando el usuario selecciona la opción correspondiente en su dispositivo móvil. Desde ese momento, cada dos centésimas de segundo se envía por internet un paquete de datos al servidor, conteniendo información sobre la situación actual en la que se encuentra el aparato según la actividad que se esté realizando. El servidor analiza cada uno de estos paquetes, los procesa y se los pasa a la red neuronal. Ésta realiza una predicción, con cierto grado de seguridad, que se almacena en la base de datos del servidor. Además de realizar predicciones, la aplicación móvil permite también enviar datos con una clase asociada relativa a la actividad que se está realizando. Estos datos se almacenan también en la base de datos y se pueden utilizar posteriormente para reentrenar la red y realizar nuevas pruebas.
Al acceder a la aplicación web a través de una URL desde un navegador, se puede abrir una página web donde se muestra, en tiempo real, la actividad que la red predice que está realizando actualmente el usuario que lleva el dispositivo móvil, junto con el porcentaje de acierto respecto de la predicción realizada.
Se describen las fases de desarrollo de la aplicación comenzando por el estudio y análisis de algoritmos de deep-learning a este ámbito de reconocimiento de actividades humanas, así como todo el proceso de ensayo y pruebas realizado sobre distintos modelos de redes para obtener los resultados más óptimos, con resultados satisfactorios.
This work explains the design and development of a human activity recognition system (HAR) using neural networks, as well as the research work carried out over this specific field during the process. The developed system consists of several modules that use and interact through the internet, being partly hosted in the cloud. Specifically, the design of the application consists of two main modules, a smartphone application developed for Android systems and a web application deployed on a server. The Smartphone app collects data from the sensors on the mobile device (acceleration and movement sensors), sending them to the server through the internet to be analysed. The web application consists of a REST api, that uses a previously trained neural network to make predictions about the activity that its being performed based on the received data. The type of network used is a recurrent network of type long-short-term memory (LSTM), based on the works published by Matlab about sequence-to-sequence classification and prediction. This is also the tool used to develop the network model and to perform the required training processes. The prediction process starts when the user selects the corresponding option on his smartphone. From that moment, every two hundredths of a second a data packag e is sent through the internet to the server, containing information about the current situation of the device, depending on the activity being carried out. The server analyses every one of this packages, processes them, and feeds them to the neural networ k. It then makes a prediction, with a certain degree of reliability, which is stored on the server database. Besides the prediction making, the app also allows for sending data tagged with the corresponding class describing the activity currently being performed. This data is also stored in the database and can be used later to train and test the net work. When accessing the web application through a URL in a browser a webpage can be expanded displaying, in real time, the current activity predicted by the network, which is carried out by the user bearing the smartphone, is shown, alongside the certainty of the prediction made. The development phases of the application are described, starting with the study and analysis of deep-learning algorithms in this field of HAR, as well as the entire trial and testing process carried out on different network models to obtain the optimal solutions, with satisfactory results.
This work explains the design and development of a human activity recognition system (HAR) using neural networks, as well as the research work carried out over this specific field during the process. The developed system consists of several modules that use and interact through the internet, being partly hosted in the cloud. Specifically, the design of the application consists of two main modules, a smartphone application developed for Android systems and a web application deployed on a server. The Smartphone app collects data from the sensors on the mobile device (acceleration and movement sensors), sending them to the server through the internet to be analysed. The web application consists of a REST api, that uses a previously trained neural network to make predictions about the activity that its being performed based on the received data. The type of network used is a recurrent network of type long-short-term memory (LSTM), based on the works published by Matlab about sequence-to-sequence classification and prediction. This is also the tool used to develop the network model and to perform the required training processes. The prediction process starts when the user selects the corresponding option on his smartphone. From that moment, every two hundredths of a second a data packag e is sent through the internet to the server, containing information about the current situation of the device, depending on the activity being carried out. The server analyses every one of this packages, processes them, and feeds them to the neural networ k. It then makes a prediction, with a certain degree of reliability, which is stored on the server database. Besides the prediction making, the app also allows for sending data tagged with the corresponding class describing the activity currently being performed. This data is also stored in the database and can be used later to train and test the net work. When accessing the web application through a URL in a browser a webpage can be expanded displaying, in real time, the current activity predicted by the network, which is carried out by the user bearing the smartphone, is shown, alongside the certainty of the prediction made. The development phases of the application are described, starting with the study and analysis of deep-learning algorithms in this field of HAR, as well as the entire trial and testing process carried out on different network models to obtain the optimal solutions, with satisfactory results.
Description
Trabajo de Fin de Grado en Ingeniería del Software, Facultad de Informática UCM, Departamento de Ingeniería de Software e Inteligencia Artificial, Curso 2019/2020