GNSS Denied Autonomous Navigation Technology Of Unmanned Aerial Vehicles In Unknown Environment

With the increasing application of unmanned aerial vehicles(UAVs)in modern society,autonomous navigation technology of UAVs has always been an important research content of UAVs application.The application scenarios and tasks of UAVs are becoming more complex and diversified,GNSS denied and unknown environments are more common in practical application.This thesis studies the GNSS denied autonomous navigation technology of UAVs in unknown environments.Unable to use prior information and GNSS positioning,UAVs need to collect information entirely by their own sensors.Due to the limited payload of UAVs,while monocular camera and IMU have advantages of low price,lightness and complementary advantages,visual inertial fusion is an ideal autonomous navigation scheme for UAVs.Aiming at autonomous navigation technology of UAVs,the following problems need to be solved: 1)feature extraction and feature matching are crucial to the real-time performance and robustness of the system,but currently the feature point detection and matching scheme usually has only unilateral advantages;2)The matching effect of feature points has a great impact on the accuracy of pose estimation in monocular vision,which brings a great challenge to the accurate pose estimation of UAV.The long-term error accumulation is also a major reason restricting the system performance.3)On the basis of visual pose estimation,how to use inertia information to achieve high-precision and absolute scale visual inertial fusion pose estimation is also a challenging work.Based on the above three problems,this thesis studies the autonomous navigation technology of UAVs,mainly including the following four parts:(1)The theoretical basis of visual navigation is expounded,which provides theoretical support for the design of navigation scheme.Compared with the classical feature detection scheme,a uniform extraction strategy was proposed based on ORB feature points.Combined with the feature matching algorithm,the improved feature point extraction scheme is verified by experiments,proving the improved feature points extraction scheme has a good distribution,rotating invariance and scale adaptability,feature extraction and feature matching cost less than 20 ms,show that the feature extraction and matching algorithm in this thesis has good robustness and good real-time performance.(2)Pose estimation technology based on monocular vision.Firstly,the correlation algorithms of pose estimation and scene reconstruction are introduced,and the bundle adjustment method is selected to estimate the camera pose.The experimental results show that the bundle adjustment method can effectively eliminate the mismatching in the matching process and improve the accuracy of pose estimation.Then,each module of the monocular system is described completely.Aiming at the problem of cumulative error,a batch BA solution based on key frame and common view relationship is proposed.Finally,the simulation experiment proves that the pose estimation scheme of monocular vision has high estimation accuracy and can eliminate the cumulative error.(3)Visual inertial fusion pose estimation technology.The IMU model and IMU preintegration are derived in detail,and IMU pre-integration is used to process IMU data.In the visual inertial fusion pose estimation scheme,IMU initialization was added to the system initialization,and the inter-frame tracking was designed into two calculation modes according to whether the sparse map was updated or not.The time continuity constraint is added to ensure the validity of IMU information.The simulation results show that the fusion pose estimation scheme can provide high precision pose estimation at absolute scale and eliminate cumulative errors.(4)A visual interface is designed for the visual inertial fusion pose estimation algorithm in this thesis,and the algorithm in this thesis is comprehensively tested in the Eu Roc dataset.Experimental results show that the proposed algorithm can achieve centimeter-level positioning in the Eu Roc dataset,with the angle error within 4° and the maximum scale error within 5%,and has a high estimation accuracy.By comparing with ROVIO algorithm,it is proved that the proposed algorithm has more advantages in estimation accuracy and stability,can effectively eliminate cumulative errors in pose estimation for a long time,and the system processing speed can reach 30 Hz,with good real-time performance.