A Bayesian Approach to Switching Linear Gaussian State-Space Models for Unsupervised Time-Series Segmentation