A new method for gesture recognition based on decoding individual finger flexions from surface EMG images

Abstract

People who have suffered a trans-radial amputation are burdened daily by tasks requiring the use of their hands. Myoelectric prosthetic hands provided some level of aid to trans-radial amputees by interpreting electromyographic (EMG) signals into hand movements. However, the actions that the prosthesis can provide are extremely limited by the inability of the device’s algorithms to translate complex myoelectric signals into dexterous motions mimicking the human hand. Improvement is required in the area of surface EMG gesture recognition since the current methods of gesture recognition do not provide sufficient classification accuracy over gesture sets that are large enough for adequate prostheses dexterity. All possible hand gesture can be described by the movements of only five individual fingers. This study will explore a method for hand gesture recognition based on decoding individual finger flexions from high-density surface EMG images. A high-density surface EMG dataset for eight subjects containing eleven gestures was used to test the proposed method against the state of the art gesture recognition method. It is shown that a multi-label convolutional network can recognize the individual finger flexions contained in a set of hand gestures and use the individual finger flexions to infer the eleven gestures to an accuracy of 87.2 %. Additionally, it is shown that eleven gestures can be recognized to an accuracy of 81.1 % when the network was trained on five gestures each containing a single finger flexion