Feature fusion of sEMG and ultrasound signals in hand gesture recognition

Multi-modal sensory fusion is believed to obtain higher accuracy in gesture recognition. Its difficulty lies in mining discriminative features and fusing features from different modalities. Surface electromyography(sEMG) and ultrasound signals are typical signal modalities in gesture recognition. It is expected that the fusion of them can take advantage of the complementarity of electrophysiological information and muscle morphology information. This paper proposed two kinds of feature fusion method. The one is concatenating the manual designed sEMG and ultrasound features, and the other is a convolutional neural network (CNN) based feature exaction and fusion method for sEMG and ultrasound signals. Eight able-bodied subjects were involved to participate in the experiments. In the experiments, four channels of sEMG and A-mode ultrasound signals corresponding to 20 gestures were collected synchronously to evaluate the proposed method. The experimental results demonstrated that the fusion sEMG-ultrasound feature always outperformed the separate sEMG or ultrasound feature regardless of the feature extraction method, and as for fusion sEMG-ultrasound feature, the CNN based method achieve a high accuracy (97.38±1.49%) in 20 gestures, which surpassed the method of concatenating the manual designed features and applying machine learning algorithm (LDA, KNN, SVM).