Human Activity Recognition Using Accelerometer, Gyroscope and Magnetometer Sensors: Deep Neural Network Approaches

Abstract

Nowadays, smartphone with various extraordinary and notable sensors makes new invigorating open entryways for Data Mining and Machine Learning; other than makes another exploration field for Human Activity Recognition a.k.a. HAR. Such development advanced the manufacture of an assortment of useful datasets, which encourages the investigation of certainties of different research area. The MHEALTH, a multivariate time series dataset is such an optional dataset that was arranged in order to encourage the investigation with respect to HAR. The dataset includes data with respect to twelve human activities oppressing ten volunteers. Sensors were utilized for the information amassing process to be specific Accelerometer, Gyroscope, Magnetometer and Electrocardiogram signal. This paper plays out a relative report on HAR process as far as work of four unique information preprocessing strategies joined by six popular classifiers entitled as Random Forest, Support Vector Machine, Naive Bayes, along with three deep learning approaches, Multilayer Perceptron, Deep Convolutional Neural Network and Long-Short Term Memory. Information preprocessing strategies that were utilized on the referenced dataset are disposal of invalid name occurrences, consistency of unequal classes, low-pass filtering and Principle Component Analysis. The objective of this examination is to break down execution of various classifiers as far as the referenced information preprocessing techniques and furthermore distinguishing the procedure for which the classifiers display prevalent precision.