Towards an electromyographic armband: an embedded machine learning algorithms comparison

Gesture recognition is a trending topic in modern technology, being used to control mobile apps, robotics and also videogames. Many approaches are in use to detect gestures and make them suitable for digital processing and machine learning classifier. ?? For this specific application, data are collected starting from the surface ElectroMyoGraphic (sEMG) signals, obtained applying non-invasive electrodes on a selected skin area. The acquisition setup makes gesture recognition also suitable for Human-Machine Interface (HMI), like prosthesis and robotic limb control.?? This thesis work offers a wide overview of the machine learning algorithms used for hand gesture recognition.?? Considering low-energy consumption as key feature, the system is based on an event-driven approach focused on the Average Threshold Crossing (ATC) information. ?? This feature is obtained averaging, in a fixed time window, the number of voltage threshold crossings by the sEMG signal, which can also be seen as an index of muscle activation. ?? The first analyzed dataset, acquired with custom made boards, has involved 25 healthy people, each one performing five movements within five repeated sessions.?? Algorithms like Neural Network (NN), K-means, Support Vector Machines (SVM), Random Forest (RF) and Gaussian Mixture Model (GMM) Naïve Bayes were tested and implemented on a microcontroller (i.e., ARM Cortex M4-F processor) for real-time applications. During the assessment, requirements like power consumption and a system latency below 300ms were taken into consideration.?? For all the algorithms the system latency was way below the 300ms; in particular: 2.56ms for neural network, 185.49 μs for random forest, 140.46 μs for GMM Naïve Bayes, 61.92 μs K-means and 54.84 ms for support vector machines. ?? Power consumption analysis has been performed on the MCU obtaining the following average values: 0.54 mW for NN, 0.5131 mW for K-means, 0.9324 mW for SVM, 0.5126 for RF and 0.3758mW for Naïve Bayes.?? Further investigations were made towards the design of an armband. Electrodes (i.e. two for signal acquisition and one for reference) placement on the forearm has been deeply analyzed and an optimal setup was reached. Bringing the acquisition channels up to seven, it was possible to increase the number of recognized gestures for a total of seven active poses plus the resting position. For these new data, a preliminary machine learning study has been conducted reaching an accuracy over 88%.