Human Activity Recognition using PCA and BiLSTM Recurrent Neural Networks

This paper presents an original approach to Human Activity Recognition (HAR) tasks based on wearable sensors data. We have trained a Bidirectional Long-Short Term Memory (BLSTM) Recurrent Neural Network (RNN) model to predict the physical activity of subjects taken from the mHealth dataset. PCA has been used to reduce dimensionality of the dataset while keeping the total variation at no less than 85%. The resulting dataset consisted of 15 features instead of features at a reduction rate of 34.7% The mhealth has 12 different activities recorded out of 10 subjects. We present a comparison of the PCA-BLSTM approach to several classical machine learning models. Results indicated that the PCA-BLSTM model has registered the highest accuracy of 97.64%. The SVM algorithm has registered the least accuracy of 64.67. We recommend the application of this procedure to other HAR datasets which have higher dimensionality and complexity of human activities.