Learning Anchor-Driven Pose Estimation For Unmanned Aerial Vehicles

In recent years,UAVs have been widely used in defense and military fields such as battlefield investigation,communication relay,fire strikes,electronic warfare,with low cost,zero casualties,high mobility,reusable and many other advantages.UAV mostly rotor-based is widely used in fields of civil use such as aerial photography,express delivery,power line inspection and other activities.Because of its wide application in the military and civilian areas,countries attach great importance to the development of UAV technology.In the UAV guided landing process,the UAV's pose is an important control index.The existing ground-based vision navigation system is only concerned about the location of the UAV in the space of the three-dimensional coordinates,without pose detection.Based on the existing guidance system of the laboratory,this paper deeply studies the UAV pose estimation method based on the learning anchor.By constructing the data set and constructing the parallel neural network,the images including UAV's anchors collected in the ground camera are estimated quickly.The PnP problem and the Kalman filter are used to estimate the pose information of the UAV,and the rapid and accurate estimation of the UAV pose is achieved.The main work and research results of this paper are as follows:(1)A multi-anchor driving pose estimation method for fixed-wing UAV during landing process guided by ground-based vision navigation system was proposed.The ROS implementation of the algorithm was also studied.A set of learning anchors is defined to represent the UAV pose.Considering the complicated landing background of the UAV,a deep learning method is developed.By using the parallel architecture of CPU and GPU,the detection of high dimensional UAV features is completed.The message transmission and data processing are completed under the ROS framework.(2)An end-to-end depth convolution neural network is constructed from binocular vision image to the UAV anchor point distribution,and a fixed wing UAV pose data set is constructed.Through multiple parallel CNN networks under the caffe framework,the accurate extraction of UAV anchors is completed.The mixed data set is obtained from the physical flight data and simulation data of the outfield.The training data enhances the generalization ability of the neural network and improves network's accuracy on the test set.(3)Two kinds of estimation methods from multiple anchors to UAV pose are designed and implemented.The accuracy of PnP and EKF methods is compared with simulation and real data.The half-physical simulation environment of Pixhawk in the loop based on Gazebo was built,and the pose estimation data under different landing conditions were analyzed.The accuracy and real-time performance of the posture estimation system were verified.