Visual-Model Based Spatial Tracking in the Presence of Occlusions

Abstract

A key criticism for using a template/visual-model based object-tracker is that they often lack robustness to partial occlusions. This results from the global nature of the algorithm as opposed to operating on local features as in feature-based methods. Nevertheless, visual-model based methods could have significant advantage as they can model more complex objects and use more of the available data than their feature-based counterparts. This paper presents a novel per-pixel occlusion-rejection algorithm for a spatial (3D) visual-model based object tracker. It detects occlusions based on violations of the brightness-constancy assumption of the optical-flow algorithm, which is an essential part of the tracker. Experimental results have clearly shown that using such an occlusion-rejection algorithm facilitates accurate tracking in the presence of partial occlusions that typically cause the object tracker to fail. This is achieved with no noticeable loss of performance; the frame-rate at which the object-tracker operates would remain the same.