An Adaptive Selection of the Scale and Orientation in Kernel Based Tracking

Abstract

A new approach to adapt the kernel scale and orientation in real-time tracking is proposed. The iterative procedure, mean shift, is the key point to find the most credible target location. Though it performs well in some bad conditions, such as camera motion, partial occlusions, and background clutters, it has limited performance on tracking the object with the changing size. In this paper, the adaptive filters were modified and integrated with the mean shift process to estimate both the object position and the matrix describing the kernel shape. The previous states on both the position and the matrix are used to predict and maintain a better approximation of the kernel scale and orientation. Some experiments prove the superior performance of our new method. It has advantages in tracking the objects changing in scale or orientation and is less prone to the background clutter and occlusions.