Filtering-based Visual Inertial Odometry For Quadrotor Systems

In recent years,quadrotor has been widely used in many fields such as aerial photography,surveying and logistics.Pose estimation plays a core role of the quadrotor navigation algorithms,and it is also a problem that is difficult to solve at present.In this paper,we design a visual inertial odometer which can provide high-precision pose for quadrotors.Firstly,this paper introduces the fundamental matrix and adopts heuristic method to automat-ically select the initialized model,which effectively improves the initialization success rate in the general scenario.The pose estimation part preserves the sparse image alignment method in SVO front-end,adding the local bundle adjustment to optimize the pose and map points,and using g2o to solve the optimization problem in the project.The mapping thread chooses to triangulate the ini-tial map points at the keyframe,and then adopts the Gauss-Newton method to update the position of map points,which solves the problem of the depth filter convergence.The results of the KITTI dataset show that the improved visual odometer can achieve very high calculation speeds without loss of accuracy.Secondly,a loosely coupled visual inertial odometer is constructed based on the EKF method,in which the visual measurement part is to improve the output of the visual odometer.In this paper,the traditional loosely coupled visual inertial odometer is improved.The IMU integral information is used to accelerate the matching of visual features and provide a motion model for the visual odometer,which accelerates the convergence of the estimation.The use of existing equipment in the laboratory to build a quadrotor platform which can be directly used to verify algorithm accuracy and calculation speed.The experimental results show that the accuracy of the algorithm depends on the noise level of the IMU,and it is more suitable for the relatively gentle motion.Finally,a tightly coupled visual inertial odometer is constructed based on the square root in-formation filtering method,and then the algorithm is validated by the open data set and the built quadrotor platform.The square root information filtering method is adopted.The whole estimation process uses QR decomposition to update the information matrix,which improves the computa-tional speed of the system.The visual front end utilizes a more robust ORB feature descriptor and uses the underlying SSE instruction set of the CPU to speed up feature processing.For the problem of inconsistency of the filtering method,the observability matrix constraint method is introduced to improve the accuracy of system estimation.The experimental results show that the proposed algorithm is suitable for low-cost sensor combinations,and it is very robust to fast,large-pose mo-tion,and its calculation speed is very fast,so it is very suitable for the pose estimation of quadrotor system.