A study on IMU-Based Human Activity Recognition Using Deep Learning and Traditional Machine Learning

摘要

Human Activity Recognition (HAR) has been an increasingly popular range to do researches which stems from the ubiquitous computing. And lately, identifying activities during daily life has become one of more and more challenges. Subsequently, more and more methods can be used in the recognition of human activities such as Support Vector Machine (SVM), Random Forests (RF) which are the representatives of Traditional Machine Learning (TML) and also some Deep Learning (DL) methods like Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN). However, neither TML nor DL is suitable for all kinds of situations and various datasets. As a result, we would like to explore more about such consequences. In this paper, we discover a discrepancy and phenomenon that different sizes of collected HAR datasets may produce influences on the effectiveness of traditional machine learning methods as well as the deep learning architectures. We conduct experiments on two kinds of different datasets USC-HAD and WISDM with the best accuracy nearly 90% in DL and 87% in TML. Due to the consequences of the experiments we give a conclusion on the individual heterogeneity problems of the HAR datasets–when dealing with the HAR datasets of small scales, the TML structures are more suitable. However, conversely, when the datasets have large amount of datasets. Specifically, DL approaches such as CNN and LSTM are more sensible choices.