Color Invariant SURF in Discriminative Object Tracking

Tracking can be seen as an online learning problem, where the focus is on discriminating object from background. From this point of view, features play a key role as the tracking accuracy depends on how well the feature distinguish object and background. Current discriminative trackers use traditional features such as intensity, RGB and full body shape features. In this paper, we propose to use color invariant SURF features in the discriminative tracking. This set of invariant features has been shown to be of increased invariance and discriminative power. The resulting tracker inherits a good discrimination between object and background while keeping advantages of the discriminative tracking framework. Experiments on a dataset of 80 videos covering a wide range of tracking circumstances show that the tracker is robust to changes in object appearance, lighting conditions and able to track objects under cluttered scenes and partial occlusion.