Human Activity Recognition Using Deep Models and Its Analysis from Domain Adaptation Perspective

Abstract

Human activity recognition (HAR) is a broad area of research which solves the problem of determining a user’s activity from a set of observations recorded on video or low-level sensors (accelerometer, gyroscope, etc.) HAR has important applications in medical care and entertainment. In this paper, we address sensor-based HAR, because it could be deployed on a smartphone and eliminates the need to use additional equipment. Using machine learning methods for HAR is common. However, such, methods are vulnerable to changes in the domain of training and test data. More specifically, a model trained on data collected by one user loses accuracy when utilised by another user, because of the domain gap (differences in devices and movement pattern results in differences in sensors’ readings.) Despite significant results achieved in HAR, it is not well-investigated from domain adaptation (DA) perspective. In this paper, we implement a CNN-LSTM based architecture along with several classical machine learning methods for HAR and conduct a series of cross-domain tests. The result of this work is a collection of statistics on the performance of our model under DA task. We believe that our findings will serve as a foundation for future research in solving DA problem for HAR.