Tactile-based control of a dexterous hand prosthesis

This thesis evaluated several aspects which are necessary to interpret the intents of an amputee in order to control a dexterous hand prosthesis. Several experiments were performed to evaluate and improve the hardware of the Tactile Bracelet, a new Human-Machine-Interface presented in [Koi15]. Also all steps involved in a pattern classification-based control system were investigated to develop a system which contains preprocessing, feature extraction and classification. Relevant information contained in the data of 320 tactile sensors of the Tactile Bracelet was extracted through ROI gradients [Cas12]. Different non-parametric classifiers were tested to classify different finger and wrist movements performed by intact subjects, as well as by an hand amputee. The classifiers tested were: the k-Nearest-Neighbor (kNN) using the Euclidean distance and the Nearest Cluster Classifier (NCC) using the Euclidean distance and also using the Mahalanobis distance. Balanced accuracy of the kNN and the NCC using the Euclidean distance applied on data sets containing finger and wrist movements was for the intact subjects as well as for the data recorded from an amputee between 72% and 74%. In case of applying the same classifiers only on wrist movements, the results were above the 95% and for the data recorded from the amputee almost 100% for both classifiers. Through this experiments, it was possible to conclude that different wrist movements can be predicted robustly through simple classifiers with high accuracy, not only on intact subjects but also on amputees