Correlation filter based visual trackers for person pursuit using a low-cost Quadrotor

Abstract

We propose two model-free visual object trackers for following targets using the low-cost quadrotor Parrot AR.Drone 2.0 at low altitudes. Our trackers employ correlation filters for short-term tracking and a redetection strategy based on tracking-learning-detection (TLD). We performed an extensive quantitative evaluation of our trackers and a wide variety of existing trackers on person pursuit sequences. The results show that our trackers outperform the existing trackers. In addition, we demonstrate the applicability of our proposed trackers in a series of flight experiments in unconstrained environments using human targets and an existing visual servoing controller.