Research On Visual Inertial Odometry Based On Point And Plane Features

Visual odometer is a method to calculate robot position and orientation by analyzing a sequence of continuous images.Traditional odometers rely on optical flow or feature points to calculate camera motion information and are not robust to complex state changes such as fast rotation.In order to overcome the above shortcomings,this thesis use of inertial measurement unit and a camera with the visual inertia of odometer(VIO)overcome depend on the instability of a single sensor,and at the same time using daily artificial environment rich in flat structure,for example,wall,the ground of corridor or street,that can further improve the estimation.The main contents and work of this thesis are as follows:(1)a more robust and accurate plane feature detection method is proposed,and plane geometry is used to assist bundle adjustment and Perspective-n-Points(PNP)tracking.After initializing the point cloud map,RANSAC is used for initial estimation of the plane,and then plane parameters are refined.Accurate plane estimation is robust to depth estimation error,so a good plane can be triangulated into more accurate space points to achieve good tracking effect.To ensure real-time updates to the planes in the VIO system’s sliding Windows,divide the planes into multiple sectors for management expansion.By screening points belonging to the plane and assuming that their perfection falls on the plane,the PNP problem is solved by plane prior fusion bundle adjustment and additional plane constraints,thus improving the stability of the solution.(2)a visual inertial odometer system based on point plane fusion and Superglue loop is constructed in this dissertation,and the system does not depend on ROS communication.The system refined gravity by detecting vertical edges during initialization,combined the improved marginalization strategy and the proposed plane distance cost constraint for bundle adjustment in sliding window optimization,and used the Superglue closed-loop mechanism in loop closure to cope with strong environmental changes.The overall scheme can use smaller optimization windows on the premise of maintaining good accuracy,which significantly reduces the computational cost of the algorithm.To verify the robustness and accuracy of the proposed VIO system,tests were performed on open data sets to a very competitive level.At the same time,this thesis also developed an application that can run a single-threaded VIO system at 20 fps on a Huawei Honor phone for AR presentation.