RT info:eu-repo/semantics/doctoralThesis
T1 Monitorización y reconocimiento de actividades de personas mayores en entornos residenciales
A1 Gómez Ramos, Raúl
A2 Universidad de Valladolid. Escuela de Doctorado
K1 Human-computer interaction
K1 Deep Learning
K1 Aprendizaje profundo
K1 Internet of Things
K1 Internet de las cosas
K1 Smart Home
K1 Hogar inteligente
K1 Neural networks
K1 Redes neuronales
K1 33 Ciencias Tecnológicas
AB This doctoral thesis presents a monitoring system for the main activities of daily living that older people can perform in their homes. This system has been developed using non-intrusive technology that does not capture personal information about the users. The techniques used to detect activities can be classified into two groups: supervised learning methods and unsupervised learning methods. The thesis begins with the analysis of information from the Milan public database by applying supervised learning methods using recurrent neural network models and different filtering and data quality improvement techniques, resulting in a system capable of real-time detection of activities performed by a user within their home with high precision.The thesis proposes the creation of a proprietary database, named SDHAR-HOME, which stores information collected from a household where two users residing simultaneously using three technological groups: a network of non-intrusive sensors, a user localization system within the home, and information provided by activity bracelets. This database collects information on a total of 18 different activities.  A supervised learning method is proposed to analyse the data from SDHAR-HOME based on a customised architecture applying three layers of different neural networks: RNN, LSTM, and GRU. A different and personalised model is generated for each user, enabling real-time activity detection. Another method for detecting ADLs is developed using CNN neural networks along with a CVV-SV validation method, which provides better results than the previously proposed methods. Finally, a neural network model is developed using transformer models and attention layers to increase the accuracy and speed of the system. In parallel, the thesis presents the development of an unsupervised learning method, which does not require a labeling stage for activities to train the model, thus generating a replicable system in a larger number of households. This method is based on the combination of two HMMs: one to filter user stays and another to provide activity information.An ambient intelligence ecosystem has been developed to provide service and functionality to the user. The system integrates a social robot, which offers several functions to the user: exercises, games, video calls, or home search and detection. The robot is responsible for providing support to the user at home, helping to combat loneliness and improve their well-being and quality of life, while also enhancing security in their daily life. The ecosystem has been tested for a duration of 2 months in a real household, allowing the detection of dangerous situations and creating personalised plans using the social robot. This system establishes a mechanism for early intervention, thus promoting and facilitating the independent living of elderly individuals, providing them with a higher quality of life and greater security in their homes.
YR 2024
FD 2024
LK https://uvadoc.uva.es/handle/10324/71572
UL https://uvadoc.uva.es/handle/10324/71572
LA spa
NO Escuela de Doctorado
DS UVaDOC
RD 24-nov-2024