Automated Aerial Docking System using Vision-Based Deep Learning

Abstract

This paper presents an automated aerial docking system for unmanned aerial vehicle (UAV). The proposed automated aerial docking system consists of two subsystems: docking mechanical system and vision-based deep learning target detection/tracking system. One of the fundamental challenges during the mid-air integration phase are locking between a leader and a follower aerial vehicles and robust target detection/tracking in the air. To confront those issues, this study not only presents the design of a robust docking mechanical system, but also proposes the effective vision-based deep learning target detection/tracking system. The design of the proposed docking mechanical system is based on bi-stable characteristic. The proposed docking mechanical system acts as a drogue by itself to secure the probe, which is attached to the follower vehicle. The proposed vision-based deep learning target detection and tracking system are developed for an onboard machine learning computer platform to install it on the unmanned aerial vehicles (UAVs). For the real-time drogue detection and tracking in the air, a deep learning based single-stage detector and point-cloud based algorithms are applied. For the performance validation, the ground test and the indoor flight test are conducted using the specially devised robot arms and the quadcopter drone.