Improved visual tracking by modified kernel smoothing with random multiplier

Robustness of single kernel based visual tracking is unsatisfactory in real cases because the similarity surface is not sharp enough near the attractive basin. The searching procedure stops before reaching the correct location of the object's model, generally, in the flat area near the peak. To solve this crux as one of the tools in real time application, a modified kernel based target representation adopting a random multiplier is applied to the original tracking algorithm. It can somehow accelerate the convergence speed of the position sequence and thus make the local shape of similarity function shaper in the neighborhood of tracked object. For rigorous inference, the theoretical derivation of formulations is also given in this paper. The satisfying results supported by experiments demonstrate the promising nature of accuracy and robustness of the proposed tracking scheme. In addition, a fewer number of iterations are needed which makes our algorithm suitable for real time application.