Variational autoencoder in analysis of motion capture data

Abstract

The aim of this study was to present the applicability of Variational Autoencoders (VAE) in the analysis and generation of human gait data, represented by multivariate sequences containing rotations of skeletal joints or 3D positions of anatomically significant body points. The proposed framework offers two main advantages: it uses data solely from healthy populations for classification (instead of the usual collation of healthy and pathological data) and employs anthropometric measures to determine generated data quality.The paper investigates the semi-supervised classification problem of anomaly detection, in which the VAE is trained based on the selected class called normal for the purpose of discriminating the remaining abnormal ones. Three challenges – the recognition of the gait performed by females and males, the distinction between gait with comfortable and increased paces, as well as the discrimination of walking and running activities – are taken into consideration.To accomplish the tasks, VAE architectures operating on preprocessed motion capture sequences containing single walking and running cycles are proposed. The results achieved are satisfactory. The areas under the receiver operating characteristic curve (ROC-AUC) for the anomaly detection challenges reach 0.862 ± 0.014, 0.843 ± 0.015, and 1.000 ± 0.000, respectively, evaluated through 30 repeated subject-disjoint splits. A robust directional asymmetry pattern was observed: models trained on populations with lower within-group variability achieve consistently higher detection performance, a finding confirmed across multiple anomaly detection methodologies including PCA, One-Class SVM, and Isolation Forest. As regards the generation of gait sequences, they are diverse and contain natural-looking movements.