| At present,UAVs mostly use GNSS and inertial information fusion navigation systems.However,the reliability of GNSS will be limited in many application scenarios.Therefore,the research on UAV navigation technology under GNSS limited conditions has important application value.Under the premise of the UAV’s limited load,the fusion of monocular vision and inertial information is a good alternative to UAV navigation and positioning under restricted GNSS conditions.In this thesis,the research on UAV visual inertial fusion technology is carried out,and the key problem of position and attitude estimation in UAV navigation is solved through the fusion of monocular vision and inertial information.This article first introduces the basic theory of UAV visual navigation technology,analyzes the key technology of UAV visual navigation under GNSS restricted conditions,and studies the UAV visual pose estimation technology.Aiming at the problem of scale uncertainty in visual pose estimation technology,further research on the UAV visual inertial fusion technology,and finally carried out experimental verification and analysis of the drone pose estimation in the simulation environment and the data set,proving that the UAV visual inertial fusion pose estimation algorithm has good accuracy and stability.This article mainly completed the following aspects: 1.First,a mathematical model is established for the UAV pose estimation problem under the restricted GNSS conditions,and the key problems to be solved by the UAV pose estimation are analyzed through the mathematical model.Then analyze the key technologies in UAV pose estimation,the relationship between camera imaging model and coordinate conversion is studied,the mathematical description method to describe the motion of UAV is introduced,and finally the general steps to solve the nonlinear optimization problem are analyzed.2.Secondly,the visual pose estimation technology of feature point method is studied.In order to meet the real-time requirements of UAV pose estimation,this thesis adopts ORB feature points and uses feature point screening strategy based on quad-tree algorithm.In the initial UAV motion estimation,the Fundamental matrix and Homograph matrix are calculated in parallel and the optimal strategy is selected according to the empirical formula.After the initialization is complete,the bundle adjustment method is used to estimate the UAV motion.In order to build a sparse map for UAV motion estimation,triangulation technology is used to calculate the coordinates of spatial points,and key frame technology is used to maintain the local map.3.Then for the uncertainty of monocular visual scale,based on the visual point and pose estimation of feature point method,the research on the UAV visual inertial fusion pose estimation technology is studied.The IMU measurement model is analyzed,and the IMU Pre-integration technology is used to process the IMU data.Initialization uses a visual inertial loose coupling strategy to calculate the relative pose of the camera and IMU,calibrate the gyro’s zero offset,and complete the initialization of gravity direction,velocity,and scale factor.Then,based on the sliding window technology,the back-end optimization of the tightly coupled visual inertia is analyzed,and the solution steps of the back-end optimization are analyzed.The objective function and the Jacobin matrix of the visual error and IMU measurement error are studied,and the marginalization using the Schur complement method is used to makes it possible to maintain constant computational complexity while ensuring the computational accuracy of the system.4.Finally,the experimental and precision analysis of UAV pose estimation algorithm are carried out.Experiments of pose estimation were carried out on both the simulation environment and Eu Ro C data set respectively,and UAV dynamics and sensor simulation are carried out.Absolute pose error(APE)is used to evaluate the positioning accuracy of the pose estimation algorithm,and statistical parameters such as root mean square error are used to evaluate the absolute pose error.Experiments have proved that the visual inertial fusion pose estimation algorithm in this thesis has good accuracy,and a comparative experiment with the open source visual inertial fusion solution ROVIO proves the advance of the visual inertial fusion algorithm in accuracy and stability. |
