Vision-Only relative navigation for formation flying and aerial refueling of UAVs

Abstract

The emphasis of the thesis is placed on proposing a vision-only estimation of relative position and attitude for formation flying and aerial refueling of UAVs. It covers from image processing and state estimation. The two pose estimation algorithms were proposed for vision only estimation. One algorithm was Gaussian Least Squares Differential Correction proposed by Newton. This was known as Gauss Newton method. It guaranteed the convergence if the start point of initial value is near the true. Another approach is Levenberg-Marquadt algorithm proposed by Levenberg and Marquadt. It was interpolation of Gauss Newton method and steepest descent method. This has tuning parameter , weighting value between Gauss Newton and Steepest descent. So it was called heuristic algorithm. The accuracy and robustness of pose estimation problem are improved through some ideas. First, we reduced the noise of feature position of beacon. To do this, we applied Kalman Filter to feature position. Second, we consisted of the relative navigation filter. This filter is organized relative attitude kinematics and relative position equation. To do this we have to calculate variance covariance matrix of non-linear least squares result. Some ways are introduced in this thesis. Through this filter, we could estimate relative velocity additionally and the accuracy was improved. Two methods are adjusted to improve accuracy. And, we estimate initial state for non-linear least squares algorithm for robustness. To estimate the initial state, the object function is linearized. Vision-only estimation algorithm give target state information on image plane of the camera which can be converted into relative position and attitude information with known beacon position of leader. Actual estimation is performed in near range for validation.