Feature extraction using an AM-FM model for gait pattern classification

Abstract

This paper describes classification of gait patterns from a waist-mounted triaxial accelerometer. A feature extraction technique using empirical mode decomposition (EMD) and an amplitude/frequency modulation (AM-FM) model is proposed for the classification of walking activities from accelerometry data. A set of novel features, including AM, instantaneous frequency (IF) and instantaneous amplitude (IA), representing the walking patterns were obtained based on a second-order all-pole resonator. The back-end of the system was a 32-mixture Gaussian Mixture Model (GMM) classifier. An overall classification error rate of 4.88% was achieved for the five different human gait patterns referring to walking on flat levels, walking up and down paved ramps and walking up and down stairways.