Feature-Based Scaffolding for Object Tracking

This paper aims at the development of a video-based object tracking algorithm capable to achieve stable and reliable results in a complex situation, facing illumination variation, shape and scale change, background clutter, appearance change caused by camera moving, partial occlusions, etc. The proposal is based on modelling the target object at multiple resolutions by an attributed graph at every scale. Barycentric coordinates are an elegant way to transfer the structure information of a triangle in a particular scale to the neighbouring vertices in the graph to a different scale. The tracking is based on two steps: On the one hand a hill climbing approach is followed to track keypoints. On the other hand the structure of the graph is taken into account to filter out false assignments. The tracking process starts in a rough scale and further it is refined in lower scales to finally localize the keypoints.