Fast vision‐based autonomous detection of moving cooperative target for unmanned aerial vehicle landing. Issue 1 (17th October 2018)

• Record Type: Journal Article
• Authors: Li, Zhaoxi; Meng, Cai; Zhou, Fugen; Ding, Xilun; Wang, Xueqiang; Zhang, Huan; Guo, Pin; Meng, Xin
• Other Names: Dias Jorge guestEditor.; Lima Pedro U. guestEditor.; Seneviratne Lakmal guestEditor.; Khatib Oussama guestEditor.; Tadokoro Satoshi guestEditor.; Dario Paolo guestEditor.
• Is Part Of: Journal of field robotics. Volume 36:Issue 1(2019)
• Page Start: 34
• Page End: 48
• Publication Date: 2018-10-17
• Subjects: flight strategy -- moving cooperative target -- target detection -- unmanned aerial vehicle; Robots, Industrial -- Periodicals; Automatic control -- Periodicals; 629.892
• DOI: 10.1002/rob.21815
• Languages: English
• ISSNs: 1556-4959

Abstract

Abstract: We propose a fast and effective method, fast target detection (FTD), to detect the moving cooperative target for the unmanned aerial vehicle landing, and the target is composed of double circles and a cross. The purpose of our strategy is to land on the target. The FTD method needs to detect the target at the high and low heights. At the high height, the target appears completely and stably in the camera field. The FTD method can detect the circle and cross to rapidly reach the target center, named cross and circle–FTD ( C 2 − F T D ). To detect the cross, we propose a slope distance equation to obtain the distance between two slopes. The proposed slopes cluster method, based on the distance equation and K ‐means, is used to determine the cross center. At the low height, the target appears incompletely and unstably. Therefore, FTD methods detect only the cross, named cross–FTD ( C 1 − F T D ). We extract the cross features ( CF s) based on line segments. Then, four CF s are combined based on graph theory. Experiments on our four datasets show that FTD has rapid speed and good performance. (Our method is implemented in C++ and is available athttps://github.com/Li-Zhaoxi/UAV-Vision-Servo .) On the Mohamed Bin Zayed International Robotics Challenge datasets made we constructed, C 2 − F T D detects the target from a 960 × 540 image approximately 20 ⁢ m s per pipeline with 82.24 % F ‐measure and tracks target approximately 6.27 ⁢ m s per pipeline with 94.39 % F ‐measure. C 1 − F T D detects centers from a 480 × 270 image at approximately 4.69 ⁢ m s per image with 86.05 % F ‐measure.