Edinburgh Napier University

  Recent years, the research of gesture recognition based on surface EMG signal has become an active topic. The conventional methods mainly focus on feature engineering, but the sEMG signal is non-stationary temporally, which makes proper feature design and selection very complicated. To tackle this problem, this paper proposes a novel dynamic gesture recognition method by employing Convolutional Neural Networks. Short-time Fourier transform (STFT) and continuous wavelet transform (CWT) are introduced to model the time-frequency features of a single channel and the relationship between different channels. Meanwhile, the performance of deep learning-based methods relies on a large amount of training data. In the context of sEMG based gesture recognition, one user cannot be expected to generate tens of thousands of examples at a time. Hence, transfer learning (TL) techniques is utilized to alleviate the data generation burden imposed on a single individual and enhance the performance of the Convolutional Neural Networks. Experiment results on the sEMG data set recorded with Myo Armband indicate that the transfer learning augmented ConvNet can achieve an accuracy of 99.41%