Real-Time Wearable-Device Based Activity recognition Using Machine Learning Methods

Abstract

Classification of activities from body-worn accelerometer data to help monitor and take care of health attracts much attention from the research community. This paper proposed to design a real-time monitoring device that can identify people's actions from the accelerometer's data worn on the waist with five activities, including lying, sitting, standing, walking, and jogging. From the collected acceleration data, it is necessary to extract suitable features for real-time classification with high performance. These features are trained with machine learning algorithms that improve the efficiency of action classification. Consequently, a decision tree algorithm was embedded in the microcontroller. This programmed waist-mounted device was connected to the monitoring system via WiFi protocol. Users could monitor activities and managed data on a computer, a website, or a smartphone. The results were optimistic when the overall accuracy for the activities dataset reached 99.3% when training and classifying the activities on the computer. When experimenting with real-time wearable devices, the overall accuracy when classifying activities decreased but was still very good, reaching over 90%.