College of Engineering
Faculty Mentor: Jacob George (Electrical Engineering, University of Utah)
The purpose of this research is to develop a joint classification-regression scheme to provide consistent, accurate and proportional myoelectric control of smart-home devices. Myoelectric control uses electrical recordings from the surface of the skin to provide intuitive and natural control of assistive devices. Most myoelectric control algorithms either classify discrete actions (press one button at a time, like a keyboard) or provide proportional (regression) control of a a few actions (slide in one direction, like a mouse). Here, we introduce a new algorithm that uses a classifier to select one from a pool of regression models each trained on a singular action. We employ a Convolutional Neural Network (CNN) to classify myoelectric input as one of three gestures: 1) rotating a dial, 2) swiping downward, 3) doing nothing (i.e., hand at rest). In addition, we train two Kalman filters (KF) for regressing the degree an individual rotates or swipes. Preliminary results show 95% classification accuracy and a root mean squared error of .175 with regression. This joint classification-regression scheme provides a unique solution to a novel problem of selecting among discrete gestures that each have proportional outputs. The joint classification-regression algorithm may provide better overall performance than state-of- the-art classification and regression algorithms for myoelectric control. The results presented here also provide one of the first early demonstrations of myoelectric control for smart-home devices.