Efficient and Robust Feature Model for Visual Tracking

Long duration visual tracking of targets is quite challenging for computer vision, because the envi-ronments may be cluttered and distracting. Illumination variations and partial occlusions are two main diffi-culties in real world visual tracking. Existing methods based on hostile appearance information cannot solve these problems effectively. This paper proposes a feature-based dynamic tracking approach that can track objects with partial occlusions and varying illumination. The method represents the tracked object by an in-variant feature model. During the tracking, a new pyramid matching algorithm was used to match the object template with the observations to determine the observation likelihood. This matching is quite efficient in calculation and the spatial constraints among these features are also embedded. Instead of complicated op-timization methods, the whole model is incorporated into a Bayesian filtering framework. The experiments on real world sequences demonstrate that the method can track objects accurately and robustly even with illu-mination variations and partial occlusions.