Visual tracking achieved by adaptive sampling from hierarchical and parallel predictions

Because the inevitable ill-posedness exists in the visual information, the brain essentially needs some prior knowledge, prediction, or hypothesis to acquire a meaningful solution. From computational point of view, visual tracking is the real-time process of statistical spatiotemporal. filtering of target states from an image stream, and incremental Bayesian computation is one of the most important devices. To make Bayesian computation of the posterior density of state variables tractable for any types of probability distribution, Particle Filters (PFs) have been often employed in the real-time vision area. In this paper, we briefly review incremental Bayesian computation and PFs for visual tracking, indicate drawbacks of PFs, and then propose our framework, in which hierarchical and parallel predictions are integrated by adaptive sampling to achieve appropriate balancing of tracking accuracy and robustness. Finally, we discuss the proposed model from the viewpoint of neuroscience.