Research On Key Techniques Of Robust High-Precision Front-End For Monocular Vision Inertial Positioning

Position is undoubtedly a crucial link in the fields of automatic drive,mobile robot,virtual reality and augmented reality.Small-scale,low-cost and easy-to-operate monocular vision odometry is a widely used position method.But the simple monocular vision can not obtain the system scale,at the same time,the robustness is low under the high dynamic condition of the carrier,while the inertial data can recover the system scale and overcome the dynamic problem.Monocular inertial odometry is the focus of research in related fields.Based on the above analysis,the thesis focuses on the key technical problems of monocular vision inertial odometer.The main research contents are monocular visual inertia joint initialization method,the on-line calibration method of high precision external parameters between camera and IMU(Inertial Measurement Unit)and the assistant tracking model of fusion of IMU data,and through the improvement and innovation of related methods and models,In order to improve the robustness,real-time performance and positioning accuracy of monocular inertial odometry,the specific research contents and contributions are as follows:(1)A fast joint initialization method for monocular vision inertia loose coupling is proposed.In order to quickly recover the scale of monocular visual inertial system,gravity and accelerometer bias,taking into account the gravity acceleration and accelerometer bias,the magnitude difference between gravity and accelerometer bias is quite different,resulting in ill-conditioned whole solution matrix and reducing the convergence of the system.So after the gravity acceleration converges,the gravity is fixed directly,at the same time,the matrix factor of the scale is reduced,and the convergence rate of the system is increased,so that the high precision and stable scale and accelerometer bias can be recovered quickly in 5 seconds to Improve the real-time performance of the system.(2)A high-precision on-line self-calibration algorithm for inertial tight coupling of monocular vision is discussed.The on-line calibration method of camera and IMU high precision external parameter is accomplished in three steps: rotation parameter calibration,translation parameter calibration and external parameter optimization model.The least square model of visual and IMU relative rotation can be constructed for the calibration of rotation parameter,and the initial rotation parameter with low precision and poor stability can be obtained.For the translation parameter calibration,the translation parameter is added to the monocular visual inertia joint initialization model as an estimate to be used as the initial state parameter of the system to be estimated.Finally,in order to improve the precision and stability of calibration of external parameter,the external parameter is added to the state estimation vector,and a monocular inertial tight coupling optimization model is constructed.The experimental results of open data set show that the external participation is applied to the state estimation vector.The final calibration error of rotation parameter is within 0.012 rad and that of translation parameter is less than 0.017 m.(3)An assistant tracking model for fusion of IMU data is constructed.The tracking model includes mismatching and pose tracking.Firstly,the motion relation of pixels between frames is deduced by IMU rotation pre-integration,and then the plane feature uniform velocity model is added to increase the constraint of mismatching elimination.Finally,the RANSAC method is introduced.Compared with the single RANSAC error matching elimination method,the experiments show that the proposed method has obvious advantages in accuracy and time-consuming,and it can shorten the time of RANSAC algorithm to half.At the same time,an average of 0.5 mismatched points are eliminated at the same time.For the IMU-assisted pose tracking model,in order to improve the robustness and positioning accuracy of the carrier under intense motion,the IMU priori data is integrated into the initial pose to improve the accuracy of the initial pose.Thus,the tracking robustness and positioning accuracy are improved effectively.