Algorithm evaluation for parallel detection and tracking of UAVs

This paper presents the evaluation of object detectors and trackers within a parallel software architecture to enable long distance UAV detection and tracking in real-time using a telescope-based optical system. The architecture combines computationally expensive deep learning-based object detectors with traditional object trackers to achieve a detection and tracking rate of 100 fps. Four object detectors, FRCNN, SSD, Retinanet and FCOS, are fine-tuned on a custom UAV dataset and integrated together with three trackers, Medianflow, KCF and MOSSE, into a parallel software architecture. The evaluation is conducted on a separate set of test images and videos. The combination of FRCNN and Medianflow shows the best performance in terms of intersection over union and center location offset on the video test set, enabling detection and tracking of UAVs at 100 fps.