Aerial Target Tracking And Relative Pose Estimation Based On Airborne Vision

Airborne vision plays an important role in the autonomous perception of unmanned aerial vehicle,which plays an important role in the application of target tracking,autonomous obstacle avoidance,formation flying and autonomous aerial refueling.Especially in the UAV autonomous air refueling process,the use of airborne vision to achieve the tracking flight and the relative attitude estimation,has a strong application value.At present,the most of UAV autonomous aerial refueling are based on feature point tracking and relative pose estimation,but this method is easy to lose target because of the deformation,out of sight,shelter and other factors.In this paper,starting from image preprocessing image,aerial object moving region,tracking and relative attitude estimation are studied.Target tracking algorithm bases on correlation filtering and relative attitude estimation algorithm based on manifold learning are proposed,having a certain practicality and innovation.The original image acquired by the vision sensor has a large amount of noise and redundant information.If it is applied directly,it will bring a lot of extra burden to the target tracking and the relative attitude estimation.Firstly,image preprocessing method is introduced and a target motion region detection method based on threshold segmentation is proposed.The algorithm is verified by images of aerial target.It can complete the task of noise filtering and aerial object moving region and provides a basis for the following research.Then,the performance of the target tracking algorithm is analyzed and compared,and designing an evaluation criteria.The principle of tracking algorithm kernel correlation filtering is studied.This tracking algorithm has poor robustness to scale changes,and an improvement method is put forward.Based on KCFScale,the standard data sets,video and video on air targets were verified and compared,proved the effectiveness of the algorithm.Finally,the manifold learning is introduced into the study of the relative pose estimation of the air target.The principle,advantages and disadvantages of several manifold learning algorithms are analyzed and compared.The MaGRNN algorithm based on manifold learning is proposed.Using OpenGL to build the plane's 3D model,accessing to 2D images and tagging.The plane's attitude data are set.The MaGRNN algorithm is validated on the plane's attitude data set.The verification results show that the manifold learning has great potential to be applied to the estimation of the relative pose of the air target.