SURFACE ELECTROMYOGRAM SIGNAL CLASSIFICATION FOR GESTURE RECOGNITION USING EXTREME LEARNING AND SINGULAR SPECTRUM ANALYSIS

Abstract

Surface based gesture recognition utilizes electrical signals generated by muscles during voluntary actions to recognize human gestures. Gesture recognition has gained significant attention due to its potential application in various domains, including human-computer interaction, rehabilitation and robotics. The primary objective of this thesis is to develop and evaluate algorithms for improving accuracy of sEMG based gesture recognition systems for systems with multiple degrees of freedom. Electromyogram (EMG) signals are crucial to record muscle activity. Several papers have been proposed about EMG signals and mostly machine learning techniques have been used to extract information from EMG signals. In this paper, a molecular-based feature extractor model has been presented. This architecture uses Singular Spectrum Analysis (SSA) to form sub-bands which are then subjected to number of statistical features extraction. The sub-bands are also used to generate textural features using the Local Graph Structure method described in this paper. The feature matrix generated using these methods has been reduced in dimensionality using the Neighborhood Component Analysis (NCA). Finally, an Extreme Learning Machine model for classification has been used for the classification of gestures into their respective classes. The model achieved an accuracy of >97% for 10 classes and outperformed its predecessors. EMG gesture recognition holds immense potential for revolutionizing human-machine interaction. By harnessing the electrical activity of muscles, we can create seamless interfaces that enable users to effortlessly control devices, interact with virtual environments, and improve the quality of life for individuals with motor impairments. Continued research and advancements in this field will unlock exciting opportunities for the development of intuitive and immersive human-machine interfaces.