Bayesian Online Learning on the Riemannian Manifold using A Dual Model with Applications to Video Object Tracking

Paper in proceeding, 2011

This paper proposes a new Bayesian framework-based online learning method on a Riemannian manifold for video objects. The basic idea is to consider the dynamic appearance of an object as a point moving on a nonlinear smoothing surface (Riemannian manifold), where a dual model is applied for estimating the posterior trajectory of this moving point at each time instant under the Bayesian framework. The key difference of our method is to use a set of particle manifold points generated from the same time instant for computing the Riemannian mean, rather than using a sliding window of manifold points at different times for computing a Riemannian mean in most existing Riemannian manifold tracking methods. The dual model uses two state variables for modeling the online learning process on Riemannian manifolds: one is for object appearances on Riemannian manifolds, another is for velocity vectors in tangent planes of manifolds. A particle filter is employed on the manifold to generate the posterior covariance estimate of object appearance. Next to that, we propose to use Gabor filter outputs on partitioned sub-areas of object bounding box as features, from which the covariance matrix of object appearance is formed. As an application example, the proposed online learning is employed to a Riemannian manifold object tracking scheme where tracking and online learning are performed alternatively. Experiments are performed on both visual-band videos and infrared videos, and compared with two existing manifold trackers that are most relevant. Results have shown that the proposed manifold online learning is very robust, and have helped to significantly improve the tracking performance in terms of tracking drift, tightness and accuracy of tracked boxes especially for objects with large pose changes.