Towards extending real-time EMG-based gesture recognition system

In biomedical applications, the Electromyographic (EMG) signal is used to record the electrical activity of the muscles during their contraction. EMG signal classification stands at the core of real time systems that aim to discriminate between user's movements without relying on other environmental conditions (as it is the case with gesture classification based on video). The classification of EMG signals is usually used to translate the sign language or to design computer interfaces based on gesture recognition. In this paper, we propose a continuation of our previous work, on a real-time automatic hand gesture recognition using fully connected neural networks and temporal features extracted from the EMG signals. This approach leads to a recognition of an increased number of gestures, starting from a system trained for discriminating between a small number of classes of recognized gestures. More precisely, using an innovative transfer learning method, we preserved the performance of the system (99.31% overall accuracy compared to previous 99.78%), while doubling the number of recognized gestures.