Henrik Hesse Postdoctoral Researcher
ETH Zurich Automatic Control Laboratory
Physikstrasse 3, ETL K22.2 8092 Zurich Switzerland hesseh@ethz.ch control.ee.ethz.ch
Visual Motion Tracking for Estimation of Kite Dynamics
Max Polzin, Henrik Hesse, Tony A. Wood, and Roy S. Smith Automatic Control Laboratory, ETH Zurich This work presents a novel estimation approach for an
autonomous tethered kite system. We propose an es-timator which relies on visual motion tracking of the kite position from ground-based video recording. The proposed visual tracking approach is used to assess the quality of existing estimators and is eventually applied for real-time estimation of the kite dynamics in closed-loop operation of the AWE system developed at Fach-hochschule Nordwestschweiz (FHNW). The focus in this work is on the development of a fast and reliable visual tracking algorithm that can be used for real-time estima-tion in experiments of autonomous tethered kites. The developed visual tracking algorithm combines a fast tracking algorithm with a reliable object detector to ex-ploit advantages of both methods. The implemented tracker is a dual correlation filter (DCF), first proposed in [1]. It is not suitable for long-term robust motion track-ing since erroneous tracktrack-ing occurs occasionally. These inevitable tracking failures can be tackled by resetting the tracking algorithm with assistance of a detector [2]. The latter is much more discriminating but comes with a significant computational cost. A reliability measure was therefore added to the DCF tracking algorithm in this work to provide a threshold when the detector is triggered. This improves computational efficiency of the combined visual tracking algorithm and allows reliable real-time estimation of the kite dynamics.
To demonstrate the proposed estimator, we have imple-mented the visual tracking algorithm in Matlab which is able to track objects (40x40 pixel) in over 100 frames
(1280x960 pixel) per second. For three videos of repre-sentative kite power test scenarios (sunny, cloudy and a small kite on long lines) with 23300 frames each, we achieve accurate estimates of the kite state. The figure below highlights the effect of line dynamics in state es-timation from experimental data of the two-line system at FHNW. The markers clearly demonstrate the lag intro-duced in line-angle-based estimates which is especially evident in (up-loop) curves where line tension is low.
−100 −80 −60 −40 −20 0 20 40 0 20 40 60 80 y in m z in m vision−based line−angles direction of flight
Kite path based on filtered line angle measurements against visual motion tracking. Markers illustrate the instantaneous tracked po-sitions at two different time instances.
References:
[1] Henriques J., Caseiro R., Martins P., Batista J.: High-Speed Track-ing with Kernelized Correlation Filters. IEEE Transactions on Pat-tern Analysis and Machine Intelligence, Vol. 37, No. 3, pp. 583–596 (2015)
[2] Dollar P., Appel R., Belongie S., Perona P.: Fast Feature Pyramids for Object Detection. IEEE Transactions on Pattern Analysis and Ma-chine Intelligence, Vol. 3, No. 8, pp. 1532–1545 (2014)