Multi-Object Tracking Hierarchically in Visual Data Taken From Drones

Visual understanding tasks on the drone platform have gained considerable attention recently due to the rapid development of drones. In this paper, we present a hierarchical multi-target tracker (HMTT) for visual data taken from drones. Our approach is specifically directed against sequences shot from drone's view with several stages hierarchically performed. The detector detects objects taken from different viewing angles and the detections are filtered to ensure the correctness. Moreover, we propose a method to locate the frames in the case of camera's fast move by two-norm of the homography matrix. Based on that, performance on Multi-Object Tracking is improved with the involvement of Single Object Tracking and a re-identification subnet. Our method participated in the Multi-Object Tracking Challenge (Task 4) of VisDrone2019 benchmark and achieved state-of-the-art performance.