No abstract

No abstract

During the development of control schemes for upper-limb prostheses, the selection of a classification method is the decisive factor on predicting the correct hand movements. This contribution brings forward an approach to validate and visualize the output of a chosen classifier by simulative means. Using features extracted from a collection of recorded myoelectric signals (MES), a first training set for five different classes of hand movements is produced. Subsequentially, additional MES recordings are deployed to validate the classifier. By using the output for controlling the 3D model of a prosthetic hand, the behavior of an actual prosthesis is simulated and the results of the simulation visualized. For systematic comparison and selection of different classification methods, as well as extending the number of possible motion classes, a toolbox for MATLAB TM is currently developed. By employing these tools, data from sensors combining near-infrared (NIR) spectroscopy with electromyography (EMG) can be integrated into the classification process. Our classification results show, that existing classification schemes based on EMG data can be improved significantly by adding NIR sensor data. Employing only two combined EMG-NIR sensors, five motion classes comprising full movements, including pronation and supination, can be distinguished with 100% accuracy.

During the development of control schemes for upper-limb prostheses, the selection of a classification method is the decisive factor on predicting the correct hand movements. This contribution brings forward an approach to validate and visualize the output of a chosen classifier by simulative means. Using features extracted from a collection of recorded myoelectric signals (MES), a training set for different classes of hand movements is produced and validated with additional MES recordings. Using the output of the classifier, the behavior of an actual prosthesis is simulated by controlling the 3D model of a prosthetic hand. For systematic comparison of feature sets and classification methods, a toolbox for MATLABTM has been developed. Our classification results show, that existing classification schemes based on EMG data can be improved significantly by adding NIR sensor data. Employing only two combined EMGNIR sensors, five motion classes comprising full movements, including pronation and supination, can be distinguished with 100% accuracy.

No abstract