Robust kernel design for SSD tracker

Abstract

SSD-based object tracking has shown its improved performance compared with mean-shift and many people have made further improvements based on it. However, how kernels should be designed to better cooperate with SSD metric and Newton-style iteration remains unsolved. Our work is to find out the underlying principles for SSD kernel design, which can help make the tracker more sensitive to the object movement and make the iteration have better onestep performance. We also create a new kernel called “QPeak” to demonstrate our theorems.