Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

PAS Tracker: Position-, Appearance- and Size-Aware Multi-object Tracking in Drone Videos

: Stadler, Daniel; Sommer, Lars Wilko; Beyerer, Jürgen


Bartoli, Adrien (Ed.):
Computer Vision - ECCV 2020 Workshops. Proceedings. Pt.IV : Glasgow, UK, August 23-28, 2020
Cham: Springer Nature, 2020 (Lecture Notes in Computer Science 12538)
ISBN: 978-3-030-66822-8 (Print)
ISBN: 978-3-030-66823-5 (Online)
ISBN: 978-3-030-66824-2
European Conference on Computer Vision (ECCV) <16, 2020, Online>
Computer Vision for UAVs Workshop and Challenge <2020, Online>
Conference Paper
Fraunhofer IOSB ()
Multi-Object Tracking; object detection; drone imagery

While most multi-object tracking methods based on tracking-by-detection use either spatial or appearance cues for associating detections or apply one cue after another, our proposed PAS tracker employs a novel similarity measure that combines position, appearance and size information jointly to make full use of object representations. We further extend the PAS tracker by introducing a filtering technique to remove false positive detections, particularly in crowded scenarios, and apply a camera motion compensation model to align track positions across frames. To provide high quality detections as input for the proposed tracker, the performance of eight state-of-the-art object detectors is compared on the VisDrone MOT dataset, on which the PAS tracker achieves state-of-the-art performance and ranks 3rd in the VisDrone2020 MOT challenge. In an ablation study, we demonstrate the effectiveness of the introduced tracking components and the impact of the employed detections on the tracking performance is analyzed in detail.