Representation and synthesis of behaviour using Gaussian mixtures

摘要

We propose a statistical framework for the representation and synthesis of stochastic behaviour. Highly non-linear behaviours are modelled by a Gaussian mixture over system state change and observed history. Parameters of the model are estimated during learning from a training set. Behaviour synthesis is achieved via an iterative process of model conditioning, state change selection, and state update. Experimental results demonstrate the use of the framework for the modelling and synthesis of pedestrian trajectories.