Enabling ∼100fps detection on a landing unmanned aircraft for its on-ground vision-based recovery

Zhengjiang Cao, Kuang Zhao, Qiang Fang, Weiwei Kong, Dengqing Tang, Tianjiang Hu

Abstract: In this paper, a deep learning inspired solution is proposed and developed to enable timeliness and practicality of the pre-existing ground stereo vision guidance system for flxed-wing UAVs' safe landing. Since the ground guidance prototype was restricted within applications due to its untimeliness, eventually the vision-based detection less than 15 fps (frame per second). Under such circumstances, we employ a regression based deep learning algorithm into automatic detection on the flying aircraft in the landing sequential images. The system architecture is upgraded so as to be compatible with the novel deep learning requests, and furthermore, annotated datasets are conducted to support training and testing of the regression-based learning detection algorithm. Experimental results validate that the detection attaches 100 fps or more while the localization accuracy is kept in the same level.