Research On Omnidirectional Mobile AGV Localization Algorithm Based On IMU And Visual Fusion

The area for mobile robot researches contains many sub-problems,among which the problems concerning localization and mapping are the key problems that must be solved to realize robots autonomous motion and obstacle avoidance.The outdoor movement of the robot can depend to some extent on the information provided by the Global Positioning System(GPS).But limited by the accuracy and coverage of GPS,The realization of positioning and mapping indoors often depends on Simultaneous Localization and Mapping(SLAM)algorithm.Thanks to the improvement of computing power,the research hotspots of SLAM algorithm show the following trends: from Extended Kalman filter(EKF)to non-linear optimization,From relying on a single feature to a combination of multiple features,From relying on a single sensor to multiple sensor fusion.In this paper,in view of the shortcomings of the traditional visual SLAM algorithm,combined with the visual sensor and inertial measurement unit(IMU),a newer Visual Inertial Odometry(VIO)system based on fusion of visual point and line features is researched and proposed,which improves the robustness of the SLAM algorithm and accuracy.The main work and contributions of this paper are as follows:(1)A brief introduction to the research background and progress of the SLAM algorithm,detailed introduction to the basic knowledge involved in the SLAM problem,including robot pose description and pose transformation,camera projection model and back projection model,IMU measurement model and pre-integration process,maximum posterior probability estimation theory and nonlinear optimization theory.(2)Analyzing the problem of low integration precision of IMU Euler integral and median integral methods.This article improves the traditional integration method,and apply the 4th order Runge-Kutta(RK4)integration method to the IMU pre-integration process.Experiments show that the IMU pre-integration based on RK4 method improves the integration accuracy.(3)Aiming at the problem of low robustness caused by the use of Harris Corner in the visual front end of VINS-mono,this paper takes Qtree-FAST corner points as feature points,fusing the depth information,and uses the direct method to track,taking IMU pre-integration results as the robot's initial pose.Experiments show that the use of adaptive threshold uniform FAST feature points is useful to improve the robustness and accuracy of the system.(4)In response to the problem of tracking failure in the long corridor environment caused by only using single feature information as the visual front end in VINS-mono,this paper fuses LSD line features on the basis of FAST point features,and proposes a DPL-VIO system that combines depth point line features,and improved LSD line feature triangulation method and line segment endpoint update method.The degradation problem of traditional line feature reconstruction is solved to some extends.This paper also puts forward the initialization method and optimization method of DPL-VIO system.