Ensembled HAR Approach Using Stacking on Inertial Sensors Data

In the last few years, vast growth has been monitored in the Human Activity Recognition (HAR) domain which includes applications such as surveillance, healthcare monitoring, and smart home. The reason for this widespread growth is the scope of pocket-friendly and multi-modal data collection through already present devices such as smartphones, security cameras in home and public places, and Wi-fi devices at home as well as offices. The inclusion of deep learning (DL) techniques has made the task of activity identification simple and reduced the challenges like manual feature extraction, huge dataset handling, etc. In the proposed study we have proposed three DL models with convolutional neural networks and long short-term recurrent networks. We have further proposed an enhanced performance HAR model Ens-HAR with a stacked ensemble approach. To evaluate the proposed model's performance, we have used the UCI-HAR model which is publicly available for research. We have achieved an accuracy of 92.26%, 93.35%, 90.26%, and 94.79% on CNN, LSTM, CNN-LSTM, and Ens-HAR models respectively.