About 13 million people worldwide have lost their upper limbs [1], and this is expected to increase in the coming years even further. [2] These amputations impact a patient's functionality as well as their quality of life.
In order to improve these, surface electromyographic (sEMG) controllable prostetic devices have been developed, using the electrical activity send from the brain to the muscles as control input. [3] These signals are processes after which Hand Gesture Recognition (HGR) is performed, where the movement is classified as one of the predetermined gestures, such as creating a fist, extending or pronation the hand. In recent years, deep learning has been used to improve HGR, but most models unfortunately have poor generalizability cross-day and cross-patients. These cross-day differences are mainly due to influences such as electrode placement and cross-day variability in skin conductivity and neurophysiology. The poor generalizability cross-patient is mainly due to variability between patients, such as tissue structure, skin temperature, skin formation and velocity of blood flow. [4]
This research focuses on improving the cross-day and cross-subject HGR by creating a lightweight deep learning model for real-time HGR classification, using solely forearm sEMG and applying this to a robotic hand.
