Autonomous Drone-Person Tracking and Following in Uniform Appearance Scenarios

Drone following a person has emerged as a promising technique for various surveillance applications, garnering considerable attention from researchers over the years. Despite significant advancements reported in the literature, state-of-the-art (SOTA) methods have struggled to effectively address challenges inherent in real-world scenarios, such as the presence of distractors resembling the target person, all within stringent real-time constraints. In this study, we propose a novel drone-person tracking algorithm aimed at overcoming the challenges of person tracking within a Uniform Appearance (UA) setting in real-time. Our framework integrates several components, including a face detector (RetinaFace) for person detection and localization, a face recognizer (GhostFaceNets) to identify the target person among others in the frame, a visual object tracker for continuous target tracking across frames, and a PID controller to stabilize, follow, and update the drone’s state based on the target’s state. To ensure robust and synchronized tracking in the presence of similar distractors, we evaluate nine recent SOTA trackers using two publicly available UA tracking datasets, PTUA and D-PTUAC. The extensive real-time person following experiments conducted within the UA environment demonstrate that these SOTA trackers are both applicable and robust enough to deliver satisfactory performance in tracking and following a person via drone in UA scenarios.