Recognition of composite motions based on sEMG via deep learning

Surface electromyography(sEMG) is a reliable physiological electrical signal, which represents real-time human motion intents. And the EMG-based motion recognition has the characteristics of convenient operation, non-invasion and noninterference, which has broad application prospects. This paper focuses on composite motion(including multiple actions) recognition, such as sign language motions and handwritten motions. We proposed a novel method for composite motion recognition by using deep learning. To begin with, we defined a novel data structure called sEMG image and established convolution Neural Network designed for sEMG images. In order to reduce the demand of training data, we proposed to pre-train the network by MNIST data set based on the thought of transfer learning. To verify the methods that we proposed, we acquired and preprocessed the surface EMG signals of composite motions, including handwritten number motions and sign language motions. From the results, it can be concluded that deep learning methods perform better than traditional methods, including support vector machine(SVM) and Dynamic Time Warping(DTW). Especially in different sizes handwritten number recognition experiments, the deep learning methods is still very excellent, while accuracies of traditional methods are greatly reduced. In addition, we discovered that transfer learning can help ConvNet to quickly converge and reduce the demand for data.