An Improved Real Time Compressive Tracking

Abstract

Recently, an algorithm called compressive tracking is utilized to cope with real time tracking, which uses a very sparse measurement matrix to compress samples of target and background, and then a classifier is trained to distinguish foreground and background. However, compressive tracking extracts foreground features that are limited to the current frame and does not consider the samples of previous frames during its learning procedure. This paper presents a novel enhanced algorithm, which extracts two kinds of samples of targets: current samples and previous samples. Our method use targets information from the previous and current frames to train a novel classifier which composes of two weighted sub-classifiers based on different sample bags. Experimental results demonstrate the effectiveness and robustness of our method.