A Deep Learning Filter for Visual Drone Single Object Tracking

Object tracking is one of the most important topics in computer vision. In visual drone tracking, it is an extremely challenging due to various factors, such as camera motion, partial occlusion, and full occlusion. In this paper, we propose a deep learning filter method to relieve the above problems, which is to obtain a priori position of the object at the subsequent frame and predict its trajectory to follow up the object during occlusion. Our tracker adopts the geometric transformation of the surrounding of the object to prevent the bounding box of the object lost, and it uses context information to integrate its motion trend thereby tracking the object successfully when it reappears. Experiments on the VisDrone-SOT2018 test dataset and the VisDrone-SOT2020 val dataset illustrate the effectiveness of the proposed approach.