Abstract
Mean-Shift is a computationally low-cost method of an extremal search. In an application of feature point tracking, Mean-Shift search is sometimes not able to track the feature points because the search falls into a pitfall of an incorrect local solution when their points move widely and their features change significantly. We propose a new coarse-to-fine feature point tracking algorithm. Firstly, the Kalman-filter based feature point matching corresponds coarsely the tracking point and the detected points in a wide-area. And then Mean-shift searches the feature point finely from the corresponding point in a narrow-area. Our method resolves a disadvantage while taking the advantage of Mean-Shift. We evaluated our method to track feature points for simulation movies and two kinds of sports movies. Our method tracked more feature points than other methods for a long time.
