Video object pursuit by tri-tracker with on-line learning from positive and negative candidates

Abstract

Based on chain code, an improved Hough detection method for head detection is proposed, with which moving regions of objects are determined. During tracking process, we present a tri-tracking method (tri-tracker), on-line trained by positive and negative candidates, for tracking objects. The tracker trains three support vector machines (SVMs) initialised with a small number of labelled frames and updates the classifiers in a collaborative fashion, in which, an object is represented using a local binary pattern (LBP) histogram, RGB colour histogram and pixel-pattern-based texture feature (PPBTF) histogram, respectively. Based on the probability map created by each classifier, the final probability map forms by combing three individual probability maps. And then the peak of final probability map, which we consider as the object's position, is found by mean shift. Experiments on several video sequences show the robustness and accuracy of our proposed method.