Autonomous landing of a multicopter using computer vision

Abstract

Takeoff, normal flight, and even specialized tasks such as taking pictures, are essentially solved problems in autonomous drone flight. This notably excludes landing, which typically requires a pilot because of its inherently risky nature. This project attempts to solve the problem of autonomous drone landing using computer vision and fiducial markers - and specifically does not use GPS as a primary means of navigation during landing. The system described in this thesis extends the functionality of the widely-used, open-source ArduPilot software which runs on many drones today, and which has only primitive functionality for autonomous landing. This system is implemented as a set of ROS modules which interact with ArduPilot for control. It is designed to be easily integrated into existing ArduPilot drone systems through the addition of a companion board and gimbal-mounted camera. Results are presented to evaluate the system's performance with regard to pose estimation and landing capabilities within Gazebo 9 simulator. Guidelines are provided for migration to a physical system as a gateway to future work.