The impact of sEMG feature weight on the recognition of similar grasping gesture

This paper is aimed to study the impact of sEMG feature weight on the recognition of similar grasping gesture, of which the classification performance is hindered by their alike underlying muscle activation pattern. The sEMG were collected from six forearm hand muscles (EPB, EPI, FDS, PL, MB, ED) when subjects conducted a 4-second different grasping gestures. Then empirical mode decomposition (EMD) was employed to represent the sEMG as IMF components, of which the time-domain features (ZC and WAMP) were obtained. Then PCA was used to reduce the dimension of feature set. The optimal training set of similar grasping gestures was determined by adjusting the weight of sEMG feature in the feature space. SVM classifier and Genetic-Algorithm (GA) were used to classify the patterns of the gestures. Experimental results showed that the classification rate of a gesture was improved with the increasing proportion of its sEMG feature in training set, and different gesture exhibited different sensitivity for gesture recognition. The present work suggested that the proportion of candidate gesture should not be equal, and the contribution of different gesture should be considered to optimize the training set.