Avaliação de arquiteturas de Redes Neurais Recorrentes para Reconhecimento de Atividades Humanas utilizando acelerômetros de dispositivos móveis

Abstract

Human Activity Recognition (HAR) using mobile devices makes it possible to classify an individual’s physical activities with non-intrusive and low-cost sensor data. Thus, HAR can collaborate with physiotherapeutic treatment by allowing the monitoring of activities such as walking, running, and going up or down stairs. UniMib SHAR database provides data collected from mobile device accelerometers to be a reference for evaluating machine learning algorithms for HAR. This work aims to compare different architectures for classifying the UniMib SHAR database, combining LSTM, GRU, and BLSTM models, with optimization algorithms, and regularization techniques, resulting in almost 40 different neural network models. The results show that the BLSTM and GRU networks, combined with the RMSProp and Adam optimization algorithms, and Dropout regularization method, obtained superior performance compared to the other combinations.