SeAct: Método híbrido de segmentação de fluxos de dados de sensores para ambientes de vida assistida

• Ano de defesa: 2023
• Autor(a) principal: Venceslau, Amanda Drielly Pires
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Segmentação de dados de sensores; Método híbrido de segmentação de dados; Inferência semântica; Aprendizado de máquina; Reconhecimento de atividade humana
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/73543

Population aging worldwide demands advanced tools to continuously monitor people’s activities, supporting aging and detecting potential health issues. Ambient Assisted Living (AAL) incorporates and integrates objects and people in a non-intrusive and discreet way, with solutions that deal with everything from the collection of data streams from sensors to the analysis of data for decision-making. One of the main limitations of AAL resides in the fact that data flows need to be constantly monitored through time windows or segments, whose dimensions must be adjusted according to the actions that denote ongoing activities. However, a segment may not contain events relevant to the current action, making its analysis difficult. In this context, a growing problem in sensor data segmentation is related to capturing events that the same sensor can generate, but belonging to different activities, generating ambiguity. To solve this problem, the literature addresses different methods that learn the activity pattern of an AAL resident but do not combine or process the events generated by the sensors semantically to recognize activities. In addition, no resources allow the annotation, query, or tracking of the results of the segmentation process, making it challenging to analyze segments from heterogeneous resources, whether sensors or techniques applied in segmentation. This work proposes a hybrid method for segmenting sensor data streams for AAL, called SeAct. The SeRt ontology for segments semantic annotation is also presented. Three experiments were conducted to evaluate the proposed method, adopting two public datasets. As a result, improvements in the accuracy and precision of human activity recognition were identified over existing approaches. In addition, Competence Questions were applied to validate the SeRt.