Publication: Técnicas de computación evolutiva aplicadas a la clasificación a partir de monitores de actividad física
Identifiers
Publication date
2016
Defense date
2016-06
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
Actualmente, varios factores están haciendo que el campo de reconocimiento de actividades humanas cobre una mayor importancia, como por ejemplo, la proliferación de dispositivos “wearables” que permiten monitorizar la actividad física o la tendencia de la población mundial hacia un estilo de vida cada vez más sedentario. Este estilo de vida sedentario propio de la sociedad actual se traduce en insuficiente actividad física y se considera uno de los mayores factores de riesgo para la salud, estando entre los primeros puestos de factores de riesgo de mortalidad a nivel mundial, según la OMS [11].
De esta manera, dentro del ámbito de la salud y el bienestar, gracias al avance en la miniaturización de sensores, que incluso permite su uso incorporados a la ropa de las personas, el reconocimiento automático de actividades se presenta como una solución a problemas de diversa índole, como por ejemplo, prevención de enfermedades, envejecimiento activo, monitorización remota de enfermos, además de un amplio espectro de aplicaciones en el ámbito deportivo.
Es por esto que se convierten en dispositivos de monitorización sumamente útiles en otras áreas de investigación, introduciendo el reconocimiento de actividades humanas en la computación ubicua, el entretenimiento, el registro de actividades diarias personales o el seguimiento del rendimiento deportivo o profesional.
Con la principal motivación de explorar nuevos frentes de investigación del reconocimiento de actividades, con un enfoque distinto a los planteados hasta ahora, en este trabajo se propone un sistema de reconocimiento automático de actividades que integra un algoritmo evolutivo, para la tarea de clasificación de actividades, y un enjambre de partículas, para la realización de un clustering que mejore el aprendizaje automático.
El sistema ha sido evaluado mediante validación cruzada del tipo leave-one-subject-out, para comprobar su rendimiento en situaciones de reconocimiento independiente del sujeto, obteniendo un 52,37% de acierto.
También, se ha evaluado el sistema con validación cruzada estándar de 10-folds en cada sujeto, para analizar la capacidad del sistema en casos de clasificación dependiente del sujeto, alcanzando un 98,07% de acierto.
Un resultado significativamente más positivo que el primero, que muestra que el sistema puede tender a la personalización del reconocimiento de actividades.
Además, se ha llevado a cabo la evaluación del sistema con validación cruzada estándar de 10-folds en el conjunto de todos los sujetos, con un 70,2267% de acierto, abundándose en la conclusión expuesta más arriba, de que el sistema presenta un mejor funcionamiento en situaciones de personalización del reconocimiento de actividades.
In the current time, various factors are making the field of activity recognition become more important, such as the proliferation of wearable devices that allow to monitor physical activity or global population’s tendency towards a more sendentary lifestyle. This sedentary lifestyle is turning into insufficient physical activity and is considered one of the factors with a highest risk for health, being among the leading risk factors of mortality, regarding the WHO [11]. This way, within health and wellness field, thanks to the advance in sensor miniaturization, which even allows sensor usage incorporated to people clothes, activity automatic recognition is presented as a solution to very diverse problems, such as diseases prevention, active aging, patient remote monitoring, as well as a wide range of applications in sports. For that reason, wearable sensors happen to be extremely useful monitorizing devices in other research areas, introducing human activity recognition to ubiquitous computing, entertainment industry, daily life activities logging and sportive and professional perfomance monitoring, among others. With the main motivation of exploring new research horizons, through a different approach to the previous works, in this project, an activity automatic recognition system that integrates an evolutionary algorithm, for the activity classification task, and a particle swarm, for a clustering that improves the automatic learning, is proposed. The system has been evaluated with leave-one-subject-out (LOSO) cross validation, in order to assess its performance in situations where the recognition is subject independent, obtaining an accuracy rate of 52,37%. Also, the system has been evaluated with 10-fold standard cross validation within each subject, to analyze the system’s capacity in subject dependent classification cases, reaching an accuracy rate of 98,07%. A significantly more positive result than the first one, that shows the system might tend to personalization of activity recognition. In addition, the system evaluation has been carried with 10-fold standard cross validation within the whole set of all the subjects, getting an accuracy rate of 70,2267%, which supports the conclusion presented above that the system works better in situations of personalization of the activity recognition.
In the current time, various factors are making the field of activity recognition become more important, such as the proliferation of wearable devices that allow to monitor physical activity or global population’s tendency towards a more sendentary lifestyle. This sedentary lifestyle is turning into insufficient physical activity and is considered one of the factors with a highest risk for health, being among the leading risk factors of mortality, regarding the WHO [11]. This way, within health and wellness field, thanks to the advance in sensor miniaturization, which even allows sensor usage incorporated to people clothes, activity automatic recognition is presented as a solution to very diverse problems, such as diseases prevention, active aging, patient remote monitoring, as well as a wide range of applications in sports. For that reason, wearable sensors happen to be extremely useful monitorizing devices in other research areas, introducing human activity recognition to ubiquitous computing, entertainment industry, daily life activities logging and sportive and professional perfomance monitoring, among others. With the main motivation of exploring new research horizons, through a different approach to the previous works, in this project, an activity automatic recognition system that integrates an evolutionary algorithm, for the activity classification task, and a particle swarm, for a clustering that improves the automatic learning, is proposed. The system has been evaluated with leave-one-subject-out (LOSO) cross validation, in order to assess its performance in situations where the recognition is subject independent, obtaining an accuracy rate of 52,37%. Also, the system has been evaluated with 10-fold standard cross validation within each subject, to analyze the system’s capacity in subject dependent classification cases, reaching an accuracy rate of 98,07%. A significantly more positive result than the first one, that shows the system might tend to personalization of activity recognition. In addition, the system evaluation has been carried with 10-fold standard cross validation within the whole set of all the subjects, getting an accuracy rate of 70,2267%, which supports the conclusion presented above that the system works better in situations of personalization of the activity recognition.
Description
Keywords
Inteligencia artificial, Clasificación