Examining the Efficiency of Magnetometers in Movement Classification Systems.

Magnetic sensors are used in a large variety of applications, including pattern recognition systems. In this paper, a novel a feature extraction approach is explored, which can potentially improve the classification efficiency in magnetometer-based movement classification systems. The basic idea of the approach is that the first derivative of the magnetometer data can carry more information than the raw measurements. The proposed method was tested on a dataset composed of real measurements, which were acquired using wrist-worn wireless sensor units. The applied data were collected with the help of multiple subjects for various movement classes. Different training and validation datasets were generated and tested based on the sampling frequency, the size the processing window and the feature extraction mode. Multi-Layer Perceptron (MLP) neural networks were used as classifiers. The achieved results show that this extraction mode can improve classification efficiency in most of the setups if it is used together with features extracted using raw measurements.