Mean-shift tracker with face-adjusted model

Abstract

Mean-shift algorithm shows robust performances in various object-tracking technologies including face tracking. Due to its robustness and accuracy, mean-shift algorithm is regarded as one of the best ways to apply in object-tracking technology in computer vision fields. However, it has a drawback of getting into a bottleneck state when faced with a speedy object moving beyond its window size within one image frame interval time. The time required to calculate mean-shift vector could be much lessened with lesser memory when color model is adjusted to the previously known target information. This paper shows the building process of target-adjusted model with a non-uniform quantization. The target color model dealt in this paper is the one used for deriving mean-shift vector. It is a kernel model containing both the color and distance information. This paper gives scheme to efficiently deal with color information in the model. Through a proper selection of color bins, unimportant color values were reduced to a small amount. As a result, the computing time of the mean-shift vector in face-tracking was shortened while maintaining robustness and accuracy.