Domain-Shift Manifold Online Learning and Tracking of Video Objects
Full recordShow full item record
AbstractThis paper describes a novel Grassmann manifold object tracking scheme that includes the modules of manifold online learning and occlusion handling. When objects contain significant out-of-plane pose changes, the domain where object appearances lying is shifting with time, hence a single vector space is no longer suitable for dynamic object representation.Motivated by this, we present a manifold-based scheme for tracking large out-of-plane objects (i.e. camera is close to the object) in video with online learning and long-term partial occlusion modules. The tracker uses Bayesian formulation on the manifold, performing posterior state estimation based on nonlinear state space modeling. One particle filter is applied for manifold online learning, another is for tracking. Occlusion handling is applied during the online learning to prevent learning occluding object/clutter. Tests on videos have shown very robust tracking performance when objects contain significant out-of-plane pose changes accompanied with long-term partial occlusions. Comparisons with two existing methods provide further support to the proposed method.