Finger movement detection using the electromyogram

Abstract

Upper limb amputation is a delicate condition in which people who suffer from this situation, find it extremely difficult to rehabilitate and live a comfortable life without the functionality that the fingers provide in our daily lives. The rapid development of myoelectric control would hopefully solve this kind of problem in the near future, such that advanced arm prostheses would be developed to allow these people to interact with the daily routine. This project focuses on the detection of finger movements, through the use of non-invasive surface Electromyographic (EMG) signals. These bio signals provide a direct link to muscle activations in the human body. This allows myoelectric control system to recognise the present movement being executed by skeletal muscle in the human body, or in the case of amputees, the desired movement which the human wants to perform. This clearly shows the usefulness of functional prostheses and how could these devices enhance the quality of life of a disabled person in a positive way. It was observed that for finger movement detection, there are two main approaches. The first approach is that of classifying discrete finger movements, whilst the second approach detect finger movements by estimating finger kinematics. The classification approach has been widely used by the majority of researches because it can successfully decode discrete finger movements. However, finger and hand movements are not limited to discrete movements but they also include continuous gestures. Therefore even though the finger kinematics estimation approach has been under developed, this approach permits the proportional control of a particular finger movement or gesture, which is essential for advanced arm prosthesis or a dexterous robotic hand. Therefore by using a linear regression model (LRM) and a multi-layer perceptron (MLP) neural network, finger joint angles were estimated by extracting input features from the myoelectric signals. The estimation results for both methods were compared and analysed. It was concluded that the MLP was superior to the LRM since it produced better finger kinematics estimation. The system‟s performance is assessed through the coefficient of determination (R2), the root mean square error (RMSE) and correlation coefficient (CC) measures, which for the MLP they were found to be on average, R2= 0.72, RMSE = 6.84° and CC = 0.85.