Hierarchical Classifier for Improved Human Activity Recognition using Wearable Sensors

Hierarchical classification (HC) is a multiclass classification approach that is effective in solving complex classification problems. However, HC in Human Activity Recognition (HAR) is still relatively unexplored. This paper investigates the effectiveness of HC compared to flat classification (FC) and ensemble-based approaches for classifying complex human activities based on inertial measurement unit (IMU)-based data. To conduct the comparative analysis, support vector machine (SVM) and decision tree (DT) methods are considered as the base models. and a comparison between various approaches, including traditional flat SVM, DT, SVM-based Bagging, random forest (RF), SVM-based AdaBoost, DT-based AdaBoost, SVM-based hierarchical classification (HC), and DT-based HC is carried out. Experiments conducted on two publicly available datasets, namely, mHealth and DaLiAc, indicate that HC is able to achieve better classification results than ensemble methods and is able to achieve them utilizing fewer base models.