AttnHAR: Human Activity Recognition using Data Collected from Wearable Sensors

In recent times, there has been a massive surge in demand for wearable sensing devices which accurately decode human activities. These sensors are extensively used in smartphones and smartwatches. There are a wide variety of applications of human activity recognition such as surveillance through video, healthcare, virtual reality. In this paper, we propose a hybrid deep learning architecture that learns the relation between important time points by self-attention and extracts spatio-temporal features from time-series data. The proposed approach is validated on 3 public datasets to show that self-attention enhances the predictive abilities of a neural network, namely MHEALTH, USCHAD and WISDM. We also compare the proposed model with previous works on these datasets. The result analysis show that our model performs better on these datasets achieving an overall accuracy of 95.04%, 90.91% and 99.02% respectively.