Classifying hand gestures from Surface Electromyography (sEMG) is a process which has applications in human-machine interaction, rehabilitation and prosthetic control. Reduction in the cost and increase in the availability of necessary hardware over recent years has made sEMG a more viable solution for hand gesture classification. The research challenge is the development of processes to robustly and accurately predict the current gesture based on incoming sEMG data.
This thesis presents a set of methods, techniques and designs that improve upon evaluation of, and performance on, the classification problem as a whole. These are brought together to set a new baseline for the potential classification. Evaluation is improved by careful choice of metrics and design of cross-validation techniques that account for data bias caused by common experimental techniques. A landmark study is re-evaluated with these improved techniques, and it is shown that data augmentation can be used to significantly improve upon the performance using conventional classification methods.
A novel neural network architecture and supporting improvements are presented that further improve performance and is refined such that the network can achieve similar performance with many fewer parameters than competing designs. Supporting techniques such as subject adaptation and smoothing algorithms are then explored to improve overall performance and also provide more nuanced trade-offs with various aspects of performance, such as incurred latency and prediction smoothness.
A new study is presented which compares the performance potential of medical grade electrodes and a low-cost commercial alternative showing that for a modest-sized gesture set, they can compete. The data is also used to explore data labelling in experimental design and to evaluate the numerous aspects of performance that must be traded off.
