IMU-based Human Activity Recognition using Machine Learning and Deep Learning models
Saad AlKharji, Aysha Alteneiji, Kin Poon
Abstract
Research into Human Activity Recognition (HAR) with wearable sensors is attracting a lot of attention due to its wide range of applications. In this paper, we use a microcontroller with an integrated Inertial Measurement Unit (IMU) to design various Machine Learning (ML) and Deep Learning (DL) models. Five different activities were captured: walking, walking up a staircase, walking down a staircase, jumping, and falling. Both ML and DL models were trained on a variety of IMU data. A comparison was performed among the models based on their accuracy scores and confusion matrices to determine the most effective one. The proposed LSTM model achieved an accuracy score of 99% when trained with accelerometer and gyroscope data from the IMU.