Comparative analysis among feature selection of sEMG signal for hand gesture classification by armband

This work presents a comparative study between dimensionality reduction and feature selection to classification problem for six hand gestures by sEMG signal. The classified signals are wrist flexion, wrist extension, wrist flexion for the left, wrist extension to the right, forearm supination, and forearm pronation. An armband with eight channels was used to acquire the signals from 13 subjects (8 male and 5 female). Then, 29 features from time and frequency domain were extracted. Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and Support Vector Machine (SVM) were used as classifiers. Regarding the dimensionality reduction, Principal Component Analysis and LDA were applied in the signal; for feature selection, the feature combination for wrapper method step wise forward was used. The best scenario with dimensionality reduction was obtained with QDA classifier and 80 attributes from PCA, reaching accuracies of 84%. In the second scenario, with 112 attributes (8 features), a non-linear SVM (with Gaussian kernel) reached accuracies of 91%. Both methods presented similar performances among the accuracies for each class; however, dimensionality reduction approach presented less computational cost whilst has a lower accuracy compared with feature selection approach.