Dynamic Learning Rate of Template Update for Visual Target Tracking

Abstract

The trackers based on discriminative correlation filter (DCF) have achieved remarkable performance in visual target tracking in recent years. Since the targets are usually affected by various factors such as deformation, rotation, motion blur and so on, the trackers have to update the templates for tracking online. The purpose of template update is to adapt to the target changes, the magnitude of which is closely related to the motion state of the target. Actually, the learning rate of template update indicates the weight of the historical sample, and its value is fixed in most existing trackers, which will decrease the precision of the tracker or make the tracker unstable. In this study, a new dynamic learning rate method for template update is proposed for visual target tracking. The motion state of the target is defined by the difference in target center position between the frames. Then, the learning rate is adjusted dynamically according to the motion state of the target instead of the fixed value, which could achieve better performance. Experiments on the popular datasets OTB100 and UAV123 show that with the proposed dynamic learning rate for template update, the DCF-based trackers can improve tracking accuracy and obtain better tracking stability in scenarios such as fast movement and motion blur.