Reconhecimento de atividades humanas baseado na análise de fluxo contínuo de dados simbólicos

Abstract

Smartphones sensing capabilities have enabled the development of Human Activity Recognition (HAR) solutions for better understanding human behavior through computational techniques. However, these solutions have been difficult to perform in dynamic scenarios because they do not observe data evolution over time and the high consumption of computational resources, such as memory, processing and energy. This occurs because the HAR problem for smartphones has been solved through classification models generated by offline machine learning algorithms that, in this case, are limited by a data history with little information about human activities. The problem with this approach is that human activities change constantly over time and are strongly influenced by the physical environment and the user’s profile. To overcome these problems this doctoral thesis proposes a new approach to recognize human activities based on the symbolic data streaming analysis. Our approach enables the development of low-cost, scalable HAR systems capable of adapting to data change over time. In this con- text, this thesis proposes a framework called DISTAR (DIscrete STream learning for Activity Recognition), responsible for standardizing the analysis of data stream process and generation of adaptive models that observe the data evolution over time without storing a data history. The DISTAR framework uses the symbolic representation algorithms known for reducing the dimensionality and numerosity of the data. In addition, this thesis also proposes a new adaptive online algorithm, called NOHAR (NOvelty discrete data stream for Human Activity Recognition), which uses as basis the DISTAR framework. Experimental results using three databases show that NOHAR is 13 times faster compared to the state of the art and is able to reduce memory consumption by an average of 99.97.

Movimentação de Aluno em Sala de Aula: Análise Descritiva de Dados de Wearables

“…A separação de time-frames consiste em um método de segmentação utilizado em séries temporais, representando sub-grupos de dados similares [Labrador and Yejas 2013], representando a atividade realizada. Após a separação, cada segmento pode ser analisado de forma separada e sequencial [Lima et al. 2019].…”

“…O domínio de frequência permite observar o padrão das oscilações dos valores do acelerometro com o tempo. Foram selecionados: FFT (Fast Fourier Transform) e Haar Wavelet (versão da transformada de Wavelet usada em HAR [Lima et al. 2019]).…”

“…As fórmulas dos atributos podem ser encontradas nas tabelas 3.7 e 3.8 de[Lima et al. 2019].XI Congresso Brasileiro de Informática na Educação(CBIE 2022) Anais do I Workshop de Aplicações Práticas de Learning Analytics em Instituições de Ensino no Brasil(WAPLA 2022) …”