Inhomogeneously Stacked RNN for Recognizing Hand Gestures from Magnetometer Data

Hand gesture recognition systems relying on biosignal data exclusively are mandatory for a variety of applications. In general, these systems have to meet requirements such as affordability, reliability, and mobility. In general, surface electrodes are used to obtain signals caused by the contraction of underlying muscles of the forearm. These data are then used to decode hand gestures. In this work, we evaluate the possibility of replacing the electrodes by magnetometers that are cheap and can be easily implemented in mobile devices. We propose an inhomogeneously stacked recurrent neural network for classifying hand gestures given magnetometer data. The experiments reveal that the comparably small network significantly outperforms state-of-the-art hand gesture recognition systems relying on multi-modal data. Furthermore, the proposed network requires significantly shorter windows and enables a quickly responding classification system. Also, the experiments show that the performance of the proposed system does not vary much between subjects and works outstandingly for amputees.