Gesture Recognition Based on EEMD and Cosine Laplacian Eigenmap

Gesture recognition has become a research hotspot in the fields of human–computer interaction, sign language recognition, rehabilitation training, sports medicine, etc. The surface electromyography (sEMG) signal contains abundant motion intention features which can be used for gesture recognition. Existing feature engineering methods are commonly used to extract linear features while ignoring the nonlinearity and nonstationarity of sEMG and its potential spatial representation. To solve this problem, we propose a gesture recognition method based on manifold learning, which extracts the features through time-frequency domain and dimension reduction. First, the ensemble empirical mode decomposition (EEMD) is used to remove low-frequency noise in sEMG, followed by extraction of a total of four time-frequency features. Second, we adopt a cosine Laplacian eigenmap (cos-LE) to extract the inherent manifold features of feature vector, which uses cosine similarity instead of Euclidean distance to evaluate the similarity between samples. Finally, SVM is used for classification. We conduct the experiments on the NinaPro DB5 to validate the effectiveness of the proposed method. The experimental results demonstrate that the cos-LE has a better clustering effect than other dimensionality reduction (DR) methods. In the subsets of NinaPro DB5 (Exercise A, Exercise B, and Exercise C), the proposed method achieves 99.15%, 96.47%, and 94.26%, respectively, and the accuracies of 40 and 52 gestures are 92.18% and 90.82%. The successful application of cos-LE in gesture recognition based on sEMG proves the potential of manifold learning in biological signal processing.