Inertial sensor-based movement classification with dimension reduction based on feature aggregation

Abstract

In this paper, a novel feature aggregation method is proposed for dimension reduction in inertial sensor-based pattern recognition applications. The method is utilized in a movement classification algorithm, which applies measurements of wearable sensor units attached to the wrists of the subjects. The feature aggregation is realized separately for each feature type by using a linear combination of the feature values extracted for each sensor axis. The weights are computed using genetic algorithm-based optimization, which uses the classification efficiency provided by the k-nearest neighbor (kNN) method to compute the fitness value. After the weights are determined for all feature types and the aggregated inputs are computed for both the training and validation data, a multi-layer perceptron (MLP) classifier is trained, which can be used during real-time classification. The algorithm was tested with multiple datasets based on features extracted using different processing window sizes and used sensor combinations. The achieved results show that the proposed weighted aggregation provides significantly, mostly 5-6%, higher recognition rates on validation data than when equal coefficients are utilized during aggregation. The method in the case of some datasets with the same configuration gives even higher classification efficiencies than when using the features extracted for the sensor axes separately.