Probabilistic Feature Extraction from Multivariate Time Series Using Spatio-Temporal Constraints

A novel nonlinear probabilistic feature extraction method, called Spatio-Temporal Gaussian Process Latent Variable Model, is introduced to discover generalised and continuous low dimensional representation of multivariate time series data in the presence of stylistic variations. This is achieved by incorporating a new spatio-temporal constraining prior over latent spaces within the likelihood optimisation of Gaussian Process Latent Variable Models (GPLVM). As a result, the core pattern of multivariate time series is extracted, whereas a style variability is marginalised. We validate the method by qualitative comparison of different GPLVM variants with their proposed spatio-temporal versions. In addition we provide quantitative results on a classification application, i.e. view-invariant action recognition, where imposing spatiotemporal constraints is essential. Performance analysis reveals that our spatio-temporal framework outperforms the state of the art.