| The navigation system composed of Simultaneous Localization and Mapping(SLAM)algorithms and GPS enables the drone to accurately navigate in most scenes.However,its performance may decrease under light-changing and textureless conditions,where the positioning accuracy is demanded,and the GPS signal is unstable.The above situations generally occur in artificial structured environments,such as a large-scale radio telescope.There are “line features” which comes from the edges of steel structures in these scenes.If we apply “integrated features” composed of points and lines to the SLAM system,it may enable the UAV to achieve stable and accurate navigation in either scenes where point features are rich or scarce.In response to the above requirements,this paper first studied the application methods of line features in SLAM.In theory,this paper proposed a new method based on Bundle Adjustment(BA)and estimate motion using line segments.The contents include the definition of the spatial expression of line features,the definition of line segment reprojection error and the derivation of Jacobian matrix of reprojection error function on camera pose and 3D line feature coordinates.On this basis,this paper designed an integrated graph optimization model to achieve motion estimation based on point-line integrated information.Experiments showed that the accuracy of the integrated model is 11.24% higher than that of the pure point feature model.In order to achieve integrated feature SLAM on the software level,this paper focuses on the front-end software implementation.For line screening,this paper proposed a method for eliminating line segment features and designed a graph optimization model that only contains unary edge to improve the screening speed.In order to improve the running effectiveness of the software,the parameters of the feature extraction part were optimized and the overall front-end framework was improved.Compared with the initial version of the program,the calculation speed is increased by 43.6% and it can run on the embedded computer with a frequency of 4 Hz.In order to verify the performance of SLAM algorithm,the integrated feature SLAM and ORB-SLAM algorithm were run for comparison on the Kitti dataset.The results show that the former gets better accuracy in motion estimation,and the mean squared error is 66.12% less than the latter in abundant-line-feature scenes.For the comparison in the indoor scene,the latter got lost at the corners and failed;while the former completed the test successfully and built a map with segment characteristics.In order to combine the data of GPS and SLAM effectively,this paper adopted an Extended Kalman Filter(EKF)computing architecture which was improved based on ideas of “come and count” and “calculate only part of variables in filter update stage”.The improved EKF effectively reduced the matrix computing dimension and improved the operating efficiency of the navigation system.In order to verify the overall performance of UAV navigation system,this paper used the Robot Operating System(ROS)to simulate and provide reference for EKF parameter adjustment.This navigation system is mainly applied to radio telescope measurement,so the simulation and running trajectory were set according to telescope measurement trajectory.The results showed that the position estimation accuracy of EKF with vision sensor was increased by 22.03%.Finally,a six-axis UAV platform was built for flight test.The test content includes the verification of the control system and the integrated navigation system with the integrated feature SLAM as the core,and the test of the system operating conditions in the GPS-fail environment.The results show that the application of the integrated feature SLAM can make the UAV obtain reliable location information in both complex or simple scenarios,and the outdoor test accuracy is 11.05% higher than the traditional solution. |
PDF Full Text Request