|M.Sc Student||Hasson Idan|
|Subject||Acquisition and Prediction of Gestures' Surface EMG Data|
Using Sequential Deep Learning Methods
|Department||Department of Computer Science||Supervisor||Professor Alexander Bronstein|
Technology is becoming indispensable and highly integrated into nearly every aspect of life. Today's most dominant interfaces with intelligent machines relies on sophisticated hand and finger functionality. With the growing dependency on hand functionality and gesture formation, the disparity in functional ability between healthy individual and those who received partial amputations increases. As a result, muscle-computer interfaces, which translate muscle activity to computer commands, allow individuals who have suffered from transradial amputations to utilize these human-machine interfaces in order to decrease their functionality gap.
Surface electromyography (sEMG) is a common method used to measure muscle activity. In this work, we describe a method for gesture recognition using sEMG data and detail a method for massive data collection. This data also includes transitions between sequences of gestures. The sEMG data is acquired by Myo, a novel and inexpensive Bluetooth device based on dry electrodes. We propose convolutional neural network (CNN) and convolutional recurrent neural network (CRNN) architectures for gesture recognition based on the spatial and temporal characteristics of the sEMG data. We also try to examine the usability of Myo's integrated inertial data in the gesture recognition task.
Using our new database of gestures, we perform single-user evaluation of our methods, which uses train and validation data of the same person, as well as cross-users evaluation, which uses train and validation data of different users.