A Low-Power Embedded System for Real-Time EMG based Event-Driven Gesture Recognition

Gesture recognition is an important topic in modern IoT applications, being used to control mobile apps, robotics and also videogames. Many technologies are in use to detect gestures and make them suitable to digital processing and machine learning classifier. One widely used way to collect data, especially in biomedical researches, is based on surface ElectroMyoGraphic (sEMG) signals, obtained simply applying non-invasive electrodes on the skin of the area of interest. This approach makes gesture recognition suitable for Human-Machine Interface (HMI), like prosthesis and robotic limb control. This thesis work studies a possible implementation of hand gesture recognition, using a system based on the Average Threshold Crossing (ATC) event-driven feature of the forearm sEMG signals. This feature is obtained averaging on a predefined time window the events generated when a sEMG signal goes above a voltage threshold; the obtained value is an index of muscle activation. The proposed system is composed of three acquisition boards (which acquire sEMG signals and process them to obtain the ATC values) and an Apollo2 MicroController Unit (MCU) with an ARM Cortex M4F microProcessor (&#956;P). For demonstration purpose two Bluetooth 3.0 modules have been added to communicate with an Arduino-based tank able to execute performed movements. The thesis work focuses on firmware optimization on the ARM &#956;P as well as software on Matlab® environment, in order to obtain the lowest power consumption possible, with a latency suitable for real-time applications (< 300 ms). Dataset creation has involved 25 healthy people, each one performing five movements within five repeated sessions. The neural network has been trained using the holdout validation method, implemented at low computational level, to exploit ARM library capabilities. Power consumption analysis have been performed on both acquisition channel and MCU, obtained values are 0.7mW and 0.8mW, respectively; overall power consumption results in 2.9mW. Maximum latency of the classifier has been measured 8.5 ms, that added to the acquisition windows, bring to a latency of 268.5 ms from the gesture to the effective movement of the actuator. The above results make the system suitable for wearable real-time applications.