Domain-Shift Tracking: Online Learning and Visual Object Tracking on Smooth Manifolds
Paper in proceedings, 2014
This paper describes a novel domain-shift tracking scheme that includes Bayesian formulation on the Grassmann/ Riemannian manifold for tracking, and domain-shift online
object learning as well as occlusion handling on the manifold. Since out-of-plane object images do not lie in a single vector space, smoothing manifolds are more suitable tools for describing domain-shift nature of such dynamic object images. The proposed domain-shift scheme is designed for tracking large-size dynamic objects (i.e. camera is close to the object) in video that contain significant out-of-plane pose changes, and may be accompanied with long-term partial occlusions. The main features of such
domain-shift tracker include: (a) Bayesian formulation defined on a manifold instead of vector space, performing posterior state estimation on the manifold based on nonlinear state space modeling; (b) Two particle filters defined on the manifold, one for online learning, another for tracking; (c) Occlusion handling is added to the online learning process to prevent learning occluding objects/clutter. To show the variant of domain-shift trackers, two example schemes are described: one uses instantaneous data on Riemannian manifolds, another uses a sliding-window of data on Grassmann manifolds. Tests on videos from the proposed domain shift trackers have shown very robust tracking performance when large-size objects contain significant out-of-plane pose changes accompanied with long-term partial occlusions. Comparisons with three existing state-of-the-art methods provide further support to the proposed scheme.
domainshift online learning
nonlinear state space model.