Camera-Based Pose Estimation for Fixed-Wing UAVs During Cooperative Landing Maneuvers

Abstract

The autonomous landing of fixed-wing unmanned aerial vehicles (UAVs) on moving unmanned ground vehicles (UGVs) demands precise spatial synchronization of both vehicle systems. To achieve this goal, using individual GNSS receivers and additional sensors for both vehicles and having the vehicles exchange their respective 6D pose via radio communication is a straightforward approach for mutual localization. However, failure of the radio communication or individual sensors can critically affect the safety of the landing maneuver without suitable additional sensors or fallback systems. In this work, we present and compare two methods for estimating the relative 6D pose between a UGV and a UAV using camera systems attached to the UGV. The first approach uses a stereo grayscale camera setup with fiducial markers on the underside of the wings of the UAV. The second approach uses a single RGB camera and a convolutional neural network (CNN) trained on synthetic data, which is combined with a pose tracking filter to improve the update rate. Requirements for the pose estimation system are an average position error below 0.5 m and an orientation angle error below 10 deg, while the update rate should be at least 20 Hz. Different experiments with a demonstrator vehicle setup are presented, showing the performance and advantages of the respective approaches. During flight experiments with eleven flyovers, the CNN-based system was consistently able to detect the UAV pose, while the marker-based system had detections in seven cases. The CNN-based system had RMS position errors between 0.19 m and 1.36 m, and RMS yaw angle errors between 1.2 deg and 7.7 deg, which is an average improvement of 0.33 m and 5.6 deg, respectively, compared to the marker-based system. In future work, the pose estimation will be tested in the control loop with an automated UAV and UGV.