Research On Monocular VINS Using Nonlinear Optimization

In recent years,with the rapid development of computer hardware and computer vision technology,a great breakthrough has been made in Simultaneous Localization and Mapping(SLAM)of mobile robots.Since monocular cameras are small in size and easy to inst all,monocular vision SLAM technology has become the focus of research in the field of active localization of mobile robots.However,the monocular camera inherently has the problem of scale ambiguity,The system can not be used in the actual navigation application.By integrating measurement information of Inertial Measurement Unit(IMU),the monocular SLAM system can correctly recover the scale factor,while achieving higher accuracy and robustness.Based on the non-linear optimization of multi-sensor fusion technology,this dissertation improves and builds a tightly coupled VINS(Visual Inertial Navigation System)based on the non-linear optimization framework.Because the localization and mapping technology based on nonlinear optimization is essential ly a state estimation problem,this dissertation first analyzes the correlation between the existing SLAM system and improves the system observation model based on the demand of monocular VINS.This dissertation introduces the IMU measurement in the pre-integral representation,and gets the error term after the improvement of the system.In project realization,this dissertation designs and implements a complete set of monocular VINS system based on monocular ORBSLAM.Because the introduction of IMU m easurement information makes the estimation of the initial value of system state variables more complicated,an effective and practical initialization method is adopted in this dissertation.The IMU initialization and visual initialization are separately c alculated,varibables are decoupled in IMU initializ ation,so that gyro bias,gravity vector,scale factor,acceleration bias and velocity are estimated.On the other hand,this dissertation improves the three major threads,which are motion tracking,local mapping and loop closure,for the fusion of IMU information.The main contents include: Propose a motion estimation method based on joint IMU model;Modify the elimination mechanism of redundant key frames;Improve the local bundle adjustment in the local mapping thread and pose optimization and global optimization after closed-loop correction.Finally,this dissertation tests monocular VINS on the public dataset EuRoC.The results show that the initialization step of IMU successfully recovers the scale factor of the system and solves the problem of scale ambiguity.Meanwhile,the state variables can be estimated in real-time with high accuracy,and system robustness is also improved.