Técnicas de Aprendizaje Profundo para reconocimiento de acciones de movimiento mediante los sensores de aceleración de un dispositivo móvil
Loading...
Download
Official URL
Full text at PDC
Publication date
2024
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
Clasificar acciones de movimiento mediante Aprendizaje Profundo es un problema desafiante y en constante desarrollo, y es un tema de investigación en campos muy diversos como deportes y rehabilitación. En este trabajo, se presenta un tipo de redes neuronales, las LSTM (Long Short Term Memory), y se muestra una visión general de su aplicación en el reconocimiento de acciones de movimiento capturadas mediante los sensores de aceleración de un dispositivo móvil. Con esto en mente, se emplea la plataforma de programación y cálculo numérico MATLAB para desarrollar dos vertientes de investigación con el fin de abordar un espectro más amplio. En primer lugar, mediante App Designer de MATLAB se elabora la aplicación estática, la cual permite entrenar los modelos LSTM y clasifica los datos de movimiento que estén previamente almacenados en el ordenador. A continuación, se introduce la aplicación dinámica, que clasifica datos de movimiento recogidos en el momento por los sensores de aceleración del movil mediante MATLAB Mobile. También se comparan varios resultados del entrenamiento de redes LSTM para resaltar la importancia que conlleva la elección de los parámetros involucrados en dichos modelos. Por útlimo, se plantea un enfoque global acerca de la comunicación entre un dispositivo móvil, encargado de registrar acciones de movimiento, y una aplicación en la "nube", responsable de almacenar, procesar e incluso interpretar estas acciones de manera remota. Para ello se emplea una plataforma de IoT llamada ThingSpeak que esta integrada en el entorno MATLAB.
Classifying motion actions through Deep Learning is a challenging and ongoing problem, and it is the subject matter in many different fields such as sports and rehabilitation. In this work, a type of recurrent neural networks, LSTM (Long Short-term Memory), is presented followed by an overview of their application in recognizing motion actions, which are captured by the acceleration sensors of a mobile device. With this in mind, the MATLAB programming and numerical computing platform is employed to develop two lines of research, addressing a broader spectrum. First, using MATLAB’s App Designer, a static application is developed, allowing the training of neural networks and the classification of motion data previously stored on the computer. Next, a dynamic application is introduced, which classifies motion data collected in real-time by the mobile’s acceleration sensors using MATLAB Mobile. Various LSTM network training results are also compared to highlight the importance of parameter selection. Finally, a comprehensive approach to communication between a mobile device, responsible for recording motion actions, and a "cloud" application, responsible for storing, processing, and even interpreting these actions remotely, is proposed. For this purpose, an IoT (Internet of Things) platform, called ThingSpeak, integrated with MATLAB, is used.
Classifying motion actions through Deep Learning is a challenging and ongoing problem, and it is the subject matter in many different fields such as sports and rehabilitation. In this work, a type of recurrent neural networks, LSTM (Long Short-term Memory), is presented followed by an overview of their application in recognizing motion actions, which are captured by the acceleration sensors of a mobile device. With this in mind, the MATLAB programming and numerical computing platform is employed to develop two lines of research, addressing a broader spectrum. First, using MATLAB’s App Designer, a static application is developed, allowing the training of neural networks and the classification of motion data previously stored on the computer. Next, a dynamic application is introduced, which classifies motion data collected in real-time by the mobile’s acceleration sensors using MATLAB Mobile. Various LSTM network training results are also compared to highlight the importance of parameter selection. Finally, a comprehensive approach to communication between a mobile device, responsible for recording motion actions, and a "cloud" application, responsible for storing, processing, and even interpreting these actions remotely, is proposed. For this purpose, an IoT (Internet of Things) platform, called ThingSpeak, integrated with MATLAB, is used.
Description
Trabajo de Fin de Doble Grado en Ingeniería Informática y Matemáticas, Facultad de Informática UCM, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2023/2024