Comparison of time- and frequency-domain features for movement classification using data from wrist-worn sensors

Abstract

Inertial and magnetic sensors are widely used for different pattern recognition applications. In this paper, features extracted using time- and frequency-domain analysis are compared for human movement classification. Applied data were collected using wrist-mounted Wireless Sensor Network (WSN) motes equipped with 9 degree of freedom (9DOF) sensor boards. Data acquisition was done with the help of multiple subjects. To explore the capabilities of the used sensor types, different feature sets were generated and tested using multiple sensor combinations, and the feature extraction was tested utilizing raw sensor signals and computed magnitudes. The classification was done using MultiLayer Perceptron (MLP) neural networks. The obtained results show that the time-domain features (TDFs) provide higher classification efficiencies than frequency-domain features (FDFs). The highest obtained classification rate on unknown data was 91.74% using TDFs, and 88.51% applying FDFs.