Stochastic Deformable Templates and Object Tracking

Abstract

Abstract Bayesian analyses of complex structures in images using sophisticated deformable shape models are rendered possible by recent advances in Markov chain Monte Carlo methodology. We consider extending the templates into a dynamic setting to allow for object tracking in image sequences. Our method combines a well-known analytic approach to non-linear dynamic modelling, the extended Kalman filter, with MCMC algorithms designed to sample object templates from suitably-defined likelihood functions. The approach is illustrated using some real face image sequences, where the aim is to track the face and predict its next position.