Research On Autonomous Navigation Technology Of Unmanned Aerial Vehicle In Unknown Environment

With the rapid promotion of unmanned aerial vehicle(UAV),severe challenges have been posed to autonomous flight technology of unmanned aerial vehicle because of diversified flight missions and complex flight environments.In the unknown environment,which is GPS-denied or without priori geographic information,UAVs only rely on the field data collected by the airborne sensors for scene awareness and motion perception,and then perform navigation tasks and make decisions based on the perceived information.Among all kinds of airborne sensors,the camera is widely use because of its cheap cost,rich information and wide applicability.On the other hand,the navigation scheme based on computer vision technology is a research hotspot in the current navigation field,attracting a large number of researchers to carry out research on it.In this paper,this paper study the autonomous navigation and positioning technology of UAVs in unknown environment,which is based on monocular vision.Then,in order to solve the problem of scale loss causing by monocular vision,a multi-sensor fusion algorithm is designed.And this thesis verifies the algorithm through the simulation experiment.First of all,this paper analyzes the image processing,scene perception and pose estimation technology based on computer vision for UAV in the process of flight.In the stage of visual perception,the classic feature detection algorithm and tracking matching algorithm are compared.For the task requirements faced in this paper,the detection process is improved based on the FAST corner point,so that the feature points can be detected quickly and evenly.Meanwhile,Pyramid Lucas Kanade Optical Flow algorithm has been selected,which could establish matching relationships quickly and stably with the continuous observation by the camera.For the pose estimation stage,this paper analyzes and determines the mathematical model of the task of this paper.Then the method to reconstruct the scene by restoring depth information with vision technology is discussed.The pose estimation method based on two-views matching and the method of scene-based nonlinear estimation on pose are compared,which deepens our understanding of strengths and weaknesses of two methods and directs the subsequent design of the odometry system.Secondly,a real-time positioning system is designed.The monocular visual odometry is designed based on the related computer vision technology,and the specific design of each link in the system has been elaborated.This paper also use the keyframe and multi-thread technology in the implementation to optimize the real-time performance of the system.Then,in order to deal with the loss of scale information causing by monocular vision,a multi-sensor loose coupling fusion scheme is proposed based on EKF,and the pre-integration technique is used to realize the synchronization of visual and inertial measurement.The real-time positioning system,which could get the flight motion information and scene information with real scale,is finally designed based on the fusion of visual and inertial data.Finally,a simulation experiment is designed with ROS and Gazebo and its framework and simulation scene are introduced.The experiment tests the accuracy and functionality of the real-time positioning system with inertial data proposed in this paper.In the simulation experiment,the pose information and real-time flight scenes processed by the system are visualized in real time,which verifies the real-time and effectiveness of this scheme.By comparing the solution results with the real values,the experiment proves that the real-time positioning system is accuracy.