Human Activity Recognition from Accelerometer Data: Axis-Wise Versus Axes-Resultant Feature Extraction

Benefitting from the development of pervasive computing, recent years have witnessed a variety of meaningful human-centric applications, where automating the recognition of human activities plays a central role in bridging the gap between sensing data and high-level services. Accelerometer-based activity recognizer often remains a priority due to its recognition performance, low costs, and portability, however, few studies systematically investigate how to extract and use features from the time-series sensor data and further compare their discriminant power. To this end, we herein propose two different ways of extracting features and exploring their combinations. Specifically, we take as a resultant axis or separate channels the accelerometer axes and then extract axes-resultant and axis-wise features. Afterwards, we evaluate the cases where the two feature sets are used separately or jointly. Finally, we conduct comparative experiments on two public activity recognition datasets with five different classification models in terms of four performance metrics. Results show that the use of axis-wise features outperforms its competitor in the majority across the datasets and that their joint use generally leads to enhanced accuracy.