Human activity recognition using machine learning techniques in a low-resource embedded system

Human activity recognition aims to infer a person's actions from a set of observations captured by several sensors. Data acquisition, processing and inference on edge devices add a complexity factor to the task, as they involve a trade-off between hardware efficiency and performance. We present a prototype of a wearable device that identifies a person's activity: walking, running or staying still. The system consists of a Texas Instruments MSP-EXP430G2ET launchpad, connected to a BOOSTXL-SENSORS boosterpack with a BMI160 accelerometer. The designed prototype can take acceleration measurements, process them and either transmit them to a computer or classify the activity in the microcontroller. Additionally, our system has LEDs to display coloured signals according to the inferred activity in real-time. The classification algorithm is based on the calculation of statistical features (mean, standard deviation, maximum and minimum) for each accelerometer axis, the application of a dimensionality reduction algorithm (LDA, Linear Discriminant Analysis) and an SVM (Support Vector Machines) classification model.