Efficient compressive sensing tracking via mixed classifier decision

Recent years have witnessed successful use of tracking-by-detection methods, with a number of promising results being achieved. Most of these algorithms use a sliding window to collect samples and then employ these samples to train and update the classifiers. They also use an updated classifier to establish the appearance model and they take the maximum response value of the classifier as the location of the target within a fixed radius. Compressive Tracking (CT) is a novel tracking-by-detection algorithm that updates the appearance model in a compressed domain. However, the conventional CT algorithm uses a single classifier to detect the target, and if the selected region drifts, the classifier may become inaccurate. Furthermore, the CT algorithm updates the classifier parameters with a constant learning rate. Therefore, if the target is completely occluded for an extended period, the classifier will instead learn the features of the covered object and the target will ultimately be lost. To overcome these problems, we present a compressive sensing tracking algorithm using mixed classifier decision. The main improvements in our algorithm are that it adopts mixed classifiers to locate the target and it applies a dynamic learning rate to update the appearance model. An experimental comparison with state-of-the-art algorithms on eight benchmark video sequences in complicated situations shows that the proposed algorithm achieves the best performance with 12 pixels on the average center location error and 66.82% on the average overlap score.