A Visual Navigation Framework for the Aerial Recovery of UAVs

This paper proposes a visual navigation framework for the aerial recovery of unmanned aerial vehicles (UAVs). The framework is employed for pose estimation between the drogue and the probe to guide the recovery UAV to safely dock with the drogue. First, a deep learning-based detector and the proposed adaptive region of interest (AROI) tracker give the region of interest (ROI) of the drogue at high speed. Second, the centroids of the markers installed on the ring of the drogue can be extracted by image processing of the ROI, which can effectively reduce the computational overhead and the noise impact from other irrelevant areas. To improve the robustness of the framework, a marker compensation method is proposed to address situations of occlusion. After all the centroid coordinates in the image are obtained, stereo vision is employed to measure the drogue pose. Then, a Kalman filter (KF) is applied to accurately estimate the drogue pose. Finally, a ground semiphysical closed-loop experiment of the docking phase of aerial recovery is developed to verify the effectiveness of the proposed framework. The experimental results show that our framework has high accuracy, strong robustness and good real-time performance.