A Feature Extraction Method for Human Action Recognition using Body-Worn Inertial Sensors

This paper proposes a new feature extraction method named as robust linear discriminant analysis (RLDA) in human action recognition using body-worn inertial sensors. The new method is based on the classical method-linear discriminant analysis(LDA), and it can eliminate certain defect in LDA. In this paper, firstly, a popular technique of dimension reduction called principal component analysis (PCA) is used to process the data, and then the eigenvalues of within-class scatter matrix can be reestimated, from which the new projection matrix can be obtained. We use the public database called Wearable Action Recognition Database to validate our method. The experimental results can illustrate that the method of this paper is feasible and effective. Especially for classification algorithm SVM, the recognition rate can reach 99.02%. At the same time, a term called dimension reduction efficiency (DRE) is defined, which is used to evaluate two popular dimension reduction techniques including PCA and random projection(RP) in the final experiment of this paper.