| Abstract [eng] |
Human-machine interface is a rapidly growing research topic which requires new analysis methods. Electromyogram is one such way to control devices using human body. Electromyogram signal specifics, analysis and potentiality to use it to control devices using mel-frequency cepstrum coefficients (which is generally used in sound such as music and voice recognition) are presented in this work. The recording of all five hand fingers (thumb, index, middle, ring and little) movements using a logger is presented in the experimental part of this work. Various types of classifiers (discriminant analysis, nearest neighbours, artificial neural network, radial basis function, decision tree and tree ensemble) are trained withmel-frequency cepstrum coefficients extracted from finger movement recording and then recognize those using separate testing data. This is done in order to investigate which classifiers are best suited for classification using mel-frequency cepstrum of electromyogram signal. Training and testing time is also kept in mind. It is also investigated which recognition system structure (one classifier with 6 classes and two serial hierarchical classifiers) is the best for the task. Best parameters for mel-frequency cepstrum coefficient extraction are found in this work. |
