Hand kinematics estimation using non-invasive surface sensors: a linear system identification approach

Hand kinematics or joints angle estimation using sensors to control prosthetic devices is one of the growing research areas in today's biomedical engineering. The need for continuous and proportional control of movements makes the regression methods more suitable than classification approaches. In system identification like regression it is possible to find a model which tracks the behavior of an input-output relation without knowing much information about the details of the inside model (black box models). In this paper we have a comprehensive study on some well-known linear models and introduce the best one for joints angle estimation in prosthetic control. The Output Error (OE) model is introduced as the best one and RMS feature as the best feature to transform sensor signals into feature domain. For the input of the model we used both surface Electromyographic (sEMG) signals and accelerometers attached on them. As these sensors are non-invasive, there is huge interest in using them for commercial purposes. We showed that this combination results in a better performance than using each of them alone. Finally, we proposed a combinational model using different pre-processing steps to estimate all joints with good estimation and low cost with less than 10 degrees of average error. This result is comparable with state of the art methods in the literature.