Gender classification of full-body biological motion of aperiodic actions using machine learning

Abstract

Gender classification of dynamic motion using machine learning has been an active research area using mostly periodic human motion action analysis. This work proposes gender classification for aperiodic dynamic overarm throwing motion actions in an unrestricted environment. Skeletal data from motion capture recordings of throwing actions are represented using posture matrices, and pre-selected kinematic variables are extracted from the skeletal sequences to create biomechanical matrices for each throw action. The posture and biomechanical time-series matrices are converted to a feature vector describing the dynamic characteristics of each sequence. Gender classification is performed on the feature vectors using multiple machine learning models, including support vector machine, k-nearest neighbor, and decision tree with and without AdaBoost. The proposed machine learning classifier models demonstrate high level of gender recognition using aperiodic overarm throwing motion measured with motion capture data. Gender classification accuracy is shown to be further improved with the novel application of biomechanical features.