Ground View Fixed-wing UAV Tracking Dataset and Experimental Evaluation

We construct a ground view vision system based on a monocular camera and processing equipment to track a small size fixed-wing unmanned aerial vehicle (UAV). The system utilizes the captured video as input and a deep learning-based tracking method to estimate the UAV position. A training algorithm is also developed to enhance the tracking performance. To give a thorough test, we construct a ground view fixed-wing UAV tracking dataset (GFUAVT). A deep learning-based method is trained on a training dataset to improve tracking performance. 22 tracking methods (11 hand-crafted methods and 11 deep learning-based methods) are tested on the dataset. A thorough analysis is drawn to illustrate the advantages and weakness of the tracking methods. The evaluation results demonstrate that tracking the UAV is far from being solved and only a small part of video frames is successfully tracked using top ranking algorithms. The dataset will be available at https://github.com/liw11111/Ground-view-fixed-wing-UAV-tracking. a a Since the paper is under review, once the paper is accepted, we will release the dataset.