Research On Visual Inerial Odometer Method Based On Comprehensive Features Of Points And Lines

With the increasingly widespread using of SLAM(Simultaneous Localization And Mapping)technology in the field of robotics,the limitations of one single sensor have become more apparent,and multi-sensor fusion technology has become the main research direction.The visual-inertial odometry framework combines the measurement data of the camera sensor and the inertial measurement unit with high accuracy and good robustness.However,the visual processing method based only on point features is challenging in scenes such as fast motion,weak texture,weak lighting,and obvious lighting changes.Line features have good performance in the above scenes,and the three-dimensional space map constructed by line features has a structure,providing necessary semantic information for subsequent navigation.In summary,the method of visual inertial odometry based on point-line features is of great significance.The research content includes:(1)The line feature method under the VIO framework.Firstly,the method of line feature extraction and matching is researched,including introducing the LSD line feature extraction method and proposing and verifying an improved line feature matching method.Secondly,aiming at numerical stability problems during iterative optimization of the back end of the system,two different line feature parameterization methods are used,and we derive the conversion formula to establish a relationship.Finally,the method of using line features for pose initialization and map initialization is researched.(2)The tightly-coupled information fusion scheme of camera and IMU sensor based on optimization theory.Firstly,introduce the related background knowledge,including bundle adjustment method,pre-integration theory and Schur complement theory.Secondly,construct residual formulas of different information sources,study information source weight distribution methods,and derive forward propagation formula of IMU covariance matrix;aiming at the problem of system optimization scale limitation because of CPU limitation,the sliding window algorithm is used for non-linear optimization,and the sliding window marginalization method and observability are analyzed,and the sliding window optimization equation is constructed.Finally,in terms of optimization problem solving,the improved Dog-leg algorithm is proposed and experimentally verified,effectively reducing the time required.The design method of visual inertial odometry system based on point and line features is researched.This paper put forward the overall framework of the visual inertial odometer system,and design the front-end and back-end modules,including thread allocation,data management,joint initialization,marginalization strategy and map management and corresponding solutions to the problems that arise during the operation.Based on the above research,ROS is used as a standard platform for software development to develop and build a VIO system.The advantages and correctness of the introduced line features are verified through simulation experiments.By the real scene data of the Eu Ro C data set and Penn COSYVIO data set,compare with the current mainstream open source frameworks ROVIO,OKVIS and VINS-Mono to verify the superiority of the system in this paper.