Research On SLAM Technology Based Monocular Vision And IMU

SLAM(Simultaneous Localization and Mapping)is a key technology in the field of robotics,which is considered as the core and foundation for realizing the complete autonomous control and the real intelligence of the mobile robot.Because of the unique advantages of visual sensors and the continuous development of 3D computer vision algorithms,vision based SLAM methods is known as a research hotpot in recent years.However,the visual SLAM method is too dependent on the texture information of the peripheral environment,and cannot handle the absence of the scene texture and the situation of the dynamic scene.The frame rate of the visual sensor is low,so it can not handle the fast motion situation.IMU(Inertial Measurement Unit)can measure the angular velocity and acceleration of the sensor itself.It can deal with the visual failure.It has the obvious complementarity with the visual sensor,and has the potential to build a more robust SLAM system.Based on this,we propose a SLAM method based on the combination of monocular vision and IMU,which can be used for robot positioning and attitude estimation under different environments.The main contents of this article include the following parts:Firstly,this paper makes an in-depth study of SLAM system combining vision and IMU,and summarizes its research background,development and research status.The analysis of the motion model of the SLAM system and the observation model and position representation method,a mathematical model for the SLAM problem,provides a theoretical basis for the construction of the visual inertia SLAM system.Secondly,we focus on the monocular vision combined with IMU based pose estimation method.Based on multi-view geometry and IMU pre-integration theory,we proposes a visual inertial odometry(VIO)method.Its front-end uses ShiTomas feature point extraction and KLT optical flow method for feature tracking.Its back-end takes position,posture,speed and sensor bias as state variables and optimizes visual error and IMU error simultaneously in tightly coupled mode.The sliding window method is adopted to control the number of optimization variables,and the edge of the far distance frame is marginalized,so that the method is real-time.Thirdly,based on the above VIO,this paper proposes a complete scheme of visual inertial SLAM system.This scheme combines bag-of-word model based loop closure and the optimization of global pose map to further optimize the pose information of VIO output,so as to reduce the cumulative error of the system.This paper based on loop closure and pose graph optimization,in particular adds the function of relocation and map reuse to the system,and further enhances the robustness and practicability of the system.Under the existing open source monocular vision SLAM framework,the computer vision library opencv and the nonlinear optimization library are used in the robot.On the ROS of operation system,we have completed the construction of the system proposed in this paper.Finally,The experimental and precision analysis of the constructed visual inertial SLAM system is carried out.The validity of the proposed visual-inertial SLAM system is verified,and the relative pose error,absolute pose error and the time required for the system operation are analyzed.The results show that he proposed system is superior to the OKVIS method in accuracy.The least RMSE of absolute pose error is up to 0.073 m,and least RMSE of the relative pose error is 0.0026 m.And the system has good real-time performance and robustness.