Reconhecimento de atividades humanas através de um smartphone

• Ano de defesa: 2014
• Autor(a) principal: Silva, Gustavo José Barbosa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Alagoas; Brasil; Programa de Pós-Graduação em Modelagem Computacional de Conhecimento; UFAL
• Palavras-chave em Português: Interação homem-máquina; Smartphones; Aprendizado do computador; Man-machine interaction; Computer learning; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://www.repositorio.ufal.br/handle/riufal/1765

Human activity recogntion is an expanding research area (ZHANG; SAWCHUK, 2013) and aims to capture the user state and its environment using heterogeneous sensors (DAVIES; SIEWIOREK; SUKTHANKAR, 2008). Through non-intrusive activities monitoring of an individual, we might infer, for instance, if he is leading a healthy lifestyle, practicing frequently dynamic activities (like walking, running, climbing and descending stairs), or if he is leading a sedentary one, spending the majority of his time static (sitting, lying or standing). Considering the health bias, the pervasive care has signi - cant potential to increase the e ciency of health care providers, but it has as one of its main problems the automatic recognition of daily human activities (ORWAT; GRAEFE; FAULWASSER, 2008). Moreover, a report with information on the activities performed by an individual during a given period can help him have a healthier and less sedentary life. It is proposed the construction of a proof of concept that can identify, with a high accuracy, common activities being performed by the user, based on data collected from a smartphone. It is also desired the generation of alerts to the user, or to other person de ned by him, when certain customizable conditions are met. And, nally, provide a history of all activities performed, including geographical information from the user, when each activity was identi ed. Before the development itself, a data set was created with 10 volunteers who performed a pre-de ned activities circuit. Three approaches were used to generate a reliable model of machine learning: impersonal, which uses data from nine users for training and one for testing, which achieved 89.4% accuracy; personal, that is focused on a single individual, training and testing the model with di erent data from him, it had the best accuracy, 98.5%; and hybrid, which uses di erent data from all ten volunteers for training and testing, and obtained 98.16% accuracy. Once generated the model, the proposed proof of concept was developed, and then tested by a volunteer in his everyday, and overall achieved satisfactory results.