| Title |
A Study on Vision-based Landing Mark Tracking Algorithm for Delivery Drones |
| Authors |
김홍봉(Hong Bong Kim) ; 송태언(Tae Eun Song) ; 박장한(Changhan Park) ; 최병인(Byungin Choi) |
| DOI |
https://doi.org/10.5573/ieie.2019.56.1.75 |
| Keywords |
Delivery drone ; Semi-autonomous flight ; Visual tracker ; MOSSE ; DSST |
| Abstract |
This paper is concerned with visual trackers for vision-based landing of delivery drones. To date most delivery drones have depended on GNSS (Global Navigation Satellite System) like GPS (Global Positioning System) for landing. However, we believe that for security reasons the drone operator should make sure the landing spot is clear by watching the streaming video from the drone and point the final landing spot in the image. In this study assuming an operational concept in which the operator points the landing spot on the drone video as the drone approaches the landing place, a comparison study of three visual trackers (MOSSE, KCF, DSST) is described. Especially, considering a distinctive property that the landing mark size increases dramatically as a drone draws near the landing place and real-time requirement of visual tracker, we propose how to adjust parameters of the visual trackers. For the size of tracking window is fixed in MOSSE and KCF filters, the time spent for filtering is almost constant as drone approaches the landing mark. However, in case of DSST, as the maximum tracking window size parameter changes from 480 pixels to 60 pixels, the filtering speed becomes almost eight times faster. When the maximum tracking window size parameter is 120 pixels, good performance is shown in terms of tracking and filtering time. Video data obtained from flying a commercial drone are utilized for a performance evaluation and comparison study, and methods to apply the results to delivery drones are proposed. |