Research On Pose Estimation For Unmanned Aircraft Based On Computer Vision

In recent years,unmanned aircraft has been widely used in many fields such as agricultural plant protection,disaster search and rescue,film production and and lowaltitude monitoring.The autonomous navigation of unmanned aircraft is the key technology among these applications and the tranditional method to achieve autonomous navigation of unmanned aircraft is combining the inertial navigation and satellite navigation.With the rapid development of computer vision,navigation technology based on computer vision tends to mature.Under the condition of the failure of traditional navigation methods,the navigation technology based on computer vision can play a very good complementary role.The acquisition of position and posture of unmanned aircraft is the prerequisites and basis for unmanned aircraft achieving obstacle avoidance,path planning,cruise and other navigation functions.Based on above background,this dissertation studies the pose estimation of unmanned aircraft based on computer vision.As for the initial pose estimation problem of monocular vision,this dissertation proposes a method which combines optical flow tracking and two geometrical models,homorgrphy matrix and fundamental matrix,to estimate the initial pose.This method can improve the optical flow tracking + homography matrix estimation method used in SVO.By detecting the parallax of the feature points tracked by the optical flow,the initial image frames can be automatically selected,then homorgraphy matrix and fundamental matrix are calculated parallelly according to the matched features.The best geometrical model is selected by scoring and the relative motion of the initial image frames can be obtained from it.The experiment result shows that the initial posture estimation error of the improved monocular visual initial pose estimation method is smaller than the optical flow tracking + homography matrix estimation method used in SVO,and the robustness is better.As for the pose estimation problem of monocular visual odometry,this dissertation fuses the visual-inertial measurements to estimate pose of unmanned aircraft based on SVO semi-direct visual odometry.The visual estimation model and inertial estimation model are aligned by preintegration model of inertial measurement unit,the prior pose estimated from inertial measurement unit can be an initial pose used in visual odometry,which can provide a good initial value for direct method to solve relative pose iteratively,and the result of visual odometry can be used in factor graph optimization to updates the state of inertial measurement unit,thus ensures the validity of the subsequent preintegration of inertial measurement unit.The experiment result shows that the absolute translation error of the improved pose estimation method is smaller compared with SVO and OKVIS method.