Continuous Estimation of Finger Joint Angles Using Muscle Activation Inputs from Surface EMG Signals

Jimson Gelbolingo Ngeo (1151132)

Surface electromyographic (sEMG) signals are often used in many robot and rehabilitation applications because these reflect the motor intentions of users very well. However, inherent problems such as electromechanical delay (EMD) are present in such applications. Many have focused on discrete classification of hand gestures, but natural hand movements are continuous. In this thesis, we present a method to predict multiple finger joint angles from sEMG signals located in the forearm using a so-called EMG-to-Muscle Activation model that parameterizes EMD. An artificial neural network (ANN) and a Gaussian Process (GP) regressor were both evaluated and used to predict joint angles in movements involving individual and simultaneous finger flexion and extension in two subjects. With ANN, results show correlations as large as 0.92 between measured and predicted finger joint angles and an overall average root-mean-square error of 5 to 12%. Using Gaussian Process gave better prediction results specially when using few training samples. Our results also show that predictions improved when the proposed muscle activation input was used compared to using conventional filtered sEMG or time-domain based features used by related studies. Lastly, a dimensionality analysis of hand and finger movements using Principal Component Analysis (PCA) was done. Our results show that the effective dimensionality is much lower than the theoretical 20 degrees-of-freedom available on the hand. This last part may suggest the existence of motor synergies in the control of the hand.