A first approach to a taxonomy-based classification framework for hand grasps

Many solutions have been proposed to help amputated subjects regain the lost functionality. In order to interact with the outer world and objects that populate it, it is crucial for these subjects to being able to perform essential grasps. In this paper we propose a preliminary solution for the online classification of 8 basics hand grasps by considering physiological signals, namely Surface Electromyography (sEMG), exploiting a quantitative taxonomy of the considered movement. The hierarchical organization of the taxonomy allows a decomposition of the classification phase between couples of movement groups. The idea is that the closest to the roots the more hard is the classification, but on the meantime the miss-classification error is less problematic, since the two movements will be close to each other. The proposed solution is subject-independent, which means that signals from many different subjects are considered by the probabilistic framework to modelize the input signals. The information has been modeled offline by using a Gaussian Mixture Model (GMM), and then testen online on a unseen subject, by using a Gaussian-based classification. In order to be able to process the signal online, an accurate preprocessing phase is needed, in particular, we apply the Wavelet Transform (Wavelet Transform) to the Electromyography (EMG) signal. Thanks to this approach we are able to develop a robust and general solution, which can adapt quickly to new subjects, with no need of long and draining training phase. In this preliminary study we were able to reach a mean accuracy of 76.5%, reaching up to 97.29% in the higher levels.