Research On Real-time 3D Mapping For UAV With Hybrid Vision System In Large Scene

In recent years,due to the rapid development of sensor and Control Technologies,Unmanned Aerial Vehicle(UAV)has been widely used in many applications such as railway and highway construction,geological exploration,wildlife conservation,emergency monitoring of agriculture disaster and so on.At the same time,with the development of Artificial Intelligence and Information Technology,3D mapping techniques also become more and more popular.However,traditional 3D mapping techniques have the problem of complex system,high cost and transport trouble.The method of 3D mapping based on UAV has many characteristics:small system component,strong expansibility,wide perspective,flexible and convenient and so on.Thus,this paper focuses on the key technical problems of 3D mapping based on UAV,including pose estimation and point cloud map reconstruction.The major research contents of this paper are:(1)3D mapping system design.Base on the characters of 3D mapping with UAV,a mapping system using DJI Matrice-100 quad copter is built.The mapping system implement 3D mapping through a 2D laser scanner derived by a dynamixel motor.A part of point cloud is transformed and treated as the depth of image features to complete forward and downward visual odometry.Visual odometry combined with the Lidar odometry were used to form a hybrid vision system providing the initial pose estimation of UAV.(2)Multi-sensor calibration.System contains two cameras,one Lidar and IMU,it is necessary to be calibrated before data fusion.Camera has some geometry distortion caused by fabrication and assembly error which can be corrected by camera calibration.To determine the installation location of each sensor,stereo camera calibration and laser-camera joint calibration were implemented.(3)Multi-sensor data fusion algorithm based on EKF.In order to estimate the pose of UAV in GPS-denied environment,a multi-data fusion method is proposed.An Error-State based EKF algorithm is also used to improve the robustness and accuracy of pose estimation by fusing Lidar odometry,visual odometry and IMU.(4)Simulation and experiments.The 3D mapping system is tested in Gazebo simulator combining with ROS firstly.The URDF model of M100 is built based on CAD file and the physics engine.It was simulated using hector_quadrotor package which is provided by ROS community.All the functions of each sensor are completely the same as the real one in simulator.To confirm the validity of our algorithm,the system is also tested in different real scenes.