Analysis of feature correlation and feature subset on Human Activity Recognition

Abstract

The analysis of smartphone data collected using accelerometer and gyroscopic sensors to recognise human behaviours has been a crucial field of study, and it has provided solutions to numerous real-world problems in domains such as healthcare and others. The data is obtained using an accelerometer and gyroscopic sensor from a smartphone for accurate prediction of human behaviours, and a feature vector of size 561 is generated. Over this info, a collection of features is estimated. We provide an overview of these features in this article, as well as a comprehensive finding about how feature selection influences recognition of different human activities.Traditionally, the Activity Recognition Chain (ARC) built on sliding windows has dominated realistic implementations, with functionality carefully tailored for case specificity. End-to-end deep learning models, which do not distinguish between representation learning and classifier optimization, have recently gained popularity for HAR for wearables, offering "out-of-thebox" modelling with superior recognition capabilities. We revisit and explore the function of feature representations in HAR using wearables in this article. We have analysed the correlation between various features generated by the accelerometer and gyroscopic signals by the body worn sensors.We have also checked on how the selection of similar features results in the recognition of a human activity and tested
our findings on three activities.