Deep Learning Model for Upper-body Action Recognition Using Body-worn Sensors

We compare different Deep Learning proposals for upper-body human activity recognition using body-worn sensors. The evaluated models have been designed so that they need only quaternion data as input and provide results in real-time, with the aim of being potentially included on a wearable system to monitor daily activities in medical applications. Among other contributions, we propose using Fast Fourier Transform feature engineering to significantly improve recognition results and introduce a new technique for data augmentation that helps the models to learn the activities with independence of subject orientation. The proposed models were evaluated using the REALDISP dataset and overcome previous methods or achieve comparative results using 30 times less features and 110 times less training instances. The winning model achieves an accuracy of 97.7 % in a subject-wise classification and of 99.5 % in a 10-fold evaluation.