Framework for object tracking with support vector machines, structural tensor and the mean shift method

Abstract

In this paper a system is presented for object tracking based on the novel connection of the one-class SVM classifier with the mean shift tracker. An object for tracking is defined by feature vectors composed of the components of the orthogonal color space, as well as local phase and coherence components of the structural tensor which convey information on texture. The binary output of the SVM is mapped into a membership field with a proposed transformation function. Tracking is performed with the continuously adaptive mean shift method operating in the membership field. The method shows high discriminative power and fast run-time properties. © 2009 Springer-Verlag Berlin Heidelberg.