Non-Invasive Human Activity Recognition Using Deep Capsule Networks With Smart-Shoes and Smart-Watch

Abstract

In this paper, we propose an approach for automatically recognizing human activities in the non-invasive manner. We utilize heterogeneous sensors built-in a smart watch and tiny wireless accelerometers embedded inside the insole of our self-made smart shoes (e-Shoes) to acquire data for the recognition task. In addition, we use the deep capsule network models for automatically learning and representing features from the sensing data. We evaluate our approach using a dataset collected from 12 subjects wearing e-Shoes and a smart watch and performing 19 activities under multiple contexts. The results demonstrate that our approach can achieve over 78% for both precision and recall on average and highlight the potential of our approach.