Robust visual tracking based on variational auto-encoding Markov chain Monte Carlo

Abstract

In this study, we present a novel visual tracker based on the variational auto-encoding Markov chain Monte Carlo (VAE-MCMC) method. A target is tracked over time with the help of multiple geometrically related supporters whose motions correlate with those of the target. Good supporters are obtained using variational auto-encoding techniques that measure the confidence of supporters in terms of marginal probabilities. These probabilities are then used in the MCMC method to search for the best state of the target. We extend the VAE-MCMC method to a variational mixture of posteriors (VampPrior)-MCMC and hierarchical VampPrior-MCMC methods. Experimental results demonstrate that the supporters are useful for robust visual tracking and that the variational auto-encoding can accurately estimate the distribution of supporters’ states. Moreover, our proposed VAE-MCMC method quantitatively and qualitatively outperforms recent state-of-the-art tracking methods. © 2019 Elsevier Inc.