Visual Guidance of Aerial Vehicle for Target Recognition and Pose Estimation

This paper proposes a deep learning-based target pose estimation method for aerial vehicle, a robotic system combining a rotary-wing UAV and a robotic arm, which is difficult to accurately acquire the pose of the target when performing tasks in the field environment. Firstly, in the offline stage, the YOLOv5 model is improved by adding angle prediction branches to make it capable of pose estimation; secondly, the offline model is used to complete target recognition, target segmentation and pose estimation in combination with online images, and the online visual point cloud is combined with statistical filtering to complete the position estimation of targets; Finally, the effectiveness and feasibility of this visual guidance method is verified through indoor static experiments and outdoor flight experiments, and this method successfully guided the aerial vehicle to complete the torch docking task.