| Abstrakt: | The field Human Activity Recognition (HAR) based on wearable sensor data has grown considerably over the recent years. One can find many practical applications of HAR, especially in healthcare and smart-home control. In this work we present a custom hardware prototype called WaveGlove in the form of a glove with five inertial sensors. Using the prototype, we acquire two datasets with different gesture vocabularies consisting of 1000 and 10000 samples, respectively.
We implement several classification methods from Classical Machine Learning as well as Deep Learning. For evaluation we use more than 10 publicly available datasets. In addition, we propose a novel self-attention based non-recurrent neural network architecture, which on average outperforms the previously reported methods. Finally, we perform an ablation study on the acquired dataset, to demonstrate the importance of multiple sensors. We show an increase in performance when using up to three sensors, with no significant improvements with more sensors.
|
|---|