Employing Particle Filters on Riemannian Manifolds for Online Domain-Shift Object Learning and Occlusion Handling

Paper in proceeding, 2014

Visual object tracking from single cameras is often employed as the basic block in a multi-camera tracking environment, and its performance naturally a--cts the multi-camera tracking system. Online learning of object model is essential for mitigating the tracking drift for highly dynamic video objects. This paper describes a domain-shift online learning and geodesic-based occlusion handling method for enhancing the robustness of manifold object tracking, especially when a large-size object (relative to an image-size) contains signifiant out-of-plane changes along with some long-term partial occlusion. The main contributions of the domain shift online learning method include: (a) Utilizing a particle filter on the manifold for online learning; (b) Bayesian formulation on the manifold, for posterior state estimation on the manifold based on nonlinear state space modeling; (c) A geodesic-based method for occlusion handling on the manifold, for preventing learning occluding objects/ clutter. The online learning method uses covariance matrices of manifold candidate objects (or, particles) at each time instant rather than from a sliding-window of objects in the conventional case, hence possibility of fast online learning. The proposed method has been applied to Riemannian manifold tracking of video objects that contain large-size objects with significant out-of-plane changes accompanied with long-term partial occlusions. The method is tested, compared and evaluated on a range of videos, results have provided strong support to the robustness of the proposed method. Discussions on computational issue and application scenario to multi-camera environment are also included.